~mmach/netext73/mesa-haswell

lp_build_pointer_set(builder, z_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, z_value, lp_build_int_vec_type(gallivm, type), ""));

858

lp_build_pointer_set(builder, s_sample_value_store, sample_loop_state.counter, LLVMBuildBitCast(builder, s_value, lp_build_int_vec_type(gallivm, type), ""));

859

lp_build_pointer_set(builder, z_fb_store, sample_loop_state.counter, z_fb);

860

lp_build_pointer_set(builder, s_fb_store, sample_loop_state.counter, s_fb);

861

}

862

}

863

864

if (key->multisample) {

865

866

* Store the post-early Z coverage mask.

867

* Recombine the resulting coverage masks post early Z into the fragment

868

* shader execution mask.

869

870

LLVMValueRef tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");

871

tmp_s_mask_or = LLVMBuildOr(builder, tmp_s_mask_or, s_mask, "");

872

LLVMBuildStore(builder, tmp_s_mask_or, s_mask_or);

873

874

if (post_depth_coverage) {

875

LLVMValueRef mask_bit_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");

876

LLVMValueRef post_depth_mask_in = LLVMBuildLoad(builder, post_depth_sample_mask_in, "");

877

mask_bit_idx = LLVMBuildAnd(builder, s_mask, lp_build_broadcast(gallivm, int_vec_type, mask_bit_idx), "");

878

post_depth_mask_in = LLVMBuildOr(builder, post_depth_mask_in, mask_bit_idx, "");

879

LLVMBuildStore(builder, post_depth_mask_in, post_depth_sample_mask_in);

880

}

881

882

LLVMBuildStore(builder, s_mask, s_mask_ptr);

883

884

lp_build_for_loop_end(&sample_loop_state);

885

886

/* recombined all the coverage masks in the shader exec mask. */

887

tmp_s_mask_or = LLVMBuildLoad(builder, s_mask_or, "");

888

lp_build_mask_update(&mask, tmp_s_mask_or);

889

890

if (key->min_samples == 1) {

891

/* for multisample Z needs to be re interpolated at pixel center */

892

lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, NULL);

893

z = interp->pos[2];

894

lp_build_mask_update(&mask, tmp_s_mask_or);

895

}

896

} else {

897

if (post_depth_coverage) {

898

LLVMValueRef post_depth_mask_in = LLVMBuildAnd(builder, lp_build_mask_value(&mask), lp_build_const_int_vec(gallivm, type, 1), "");

899

LLVMBuildStore(builder, post_depth_mask_in, post_depth_sample_mask_in);

900

}

901

}

902

903

LLVMValueRef out_sample_mask_storage = NULL;

904

if (shader->info.base.writes_samplemask) {

905

out_sample_mask_storage = lp_build_alloca(gallivm, int_vec_type, "write_mask");

906

if (key->min_samples > 1)

907

LLVMBuildStore(builder, LLVMConstNull(int_vec_type), out_sample_mask_storage);

908

}

909

910

if (post_depth_coverage) {

911

system_values.sample_mask_in = LLVMBuildLoad(builder, post_depth_sample_mask_in, "");

912

}

913

else

914

system_values.sample_mask_in = sample_mask_in;

915

if (key->multisample && key->min_samples > 1) {

916

lp_build_for_loop_begin(&sample_loop_state, gallivm,

917

lp_build_const_int32(gallivm, 0),

918

LLVMIntULT,

919

lp_build_const_int32(gallivm, key->min_samples),

920

lp_build_const_int32(gallivm, 1));

921

922

LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");

923

s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");

924

s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");

925

s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");

926

lp_build_mask_force(&mask, s_mask);

927

lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, sample_loop_state.counter);

928

system_values.sample_id = sample_loop_state.counter;

929

system_values.sample_mask_in = LLVMBuildAnd(builder, system_values.sample_mask_in,

930

lp_build_broadcast(gallivm, int_vec_type,

931

LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "")), "");

932

} else {

933

system_values.sample_id = lp_build_const_int32(gallivm, 0);

934

935

}

936

system_values.sample_pos = sample_pos_array;

937

938

lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter, mask_store, sample_loop_state.counter);

939

940

struct lp_build_fs_llvm_iface fs_iface = {

941

.base.interp_fn = fs_interp,

942

.base.fb_fetch = fs_fb_fetch,

943

.interp = interp,

944

.loop_state = &loop_state,

945

.sample_id = system_values.sample_id,

946

.mask_store = mask_store,

947

.color_ptr_ptr = color_ptr_ptr,

948

.color_stride_ptr = color_stride_ptr,

949

.color_sample_stride_ptr = color_sample_stride_ptr,

950

.key = key,

951

};

952

953

struct lp_build_tgsi_params params;

954

memset(&params, 0, sizeof(params));

955

956

params.type = type;

957

params.mask = &mask;

958

params.fs_iface = &fs_iface.base;

959

params.consts_ptr = consts_ptr;

960

params.const_sizes_ptr = num_consts_ptr;

961

params.system_values = &system_values;

962

params.inputs = interp->inputs;

963

params.context_ptr = context_ptr;

964

params.thread_data_ptr = thread_data_ptr;

965

params.sampler = sampler;

966

params.info = &shader->info.base;

967

params.ssbo_ptr = ssbo_ptr;

968

params.ssbo_sizes_ptr = num_ssbo_ptr;

969

params.image = image;

970

params.aniso_filter_table = lp_jit_context_aniso_filter_table(gallivm, context_ptr);

971

972

/* Build the actual shader */

973

if (shader->base.type == PIPE_SHADER_IR_TGSI)

974

lp_build_tgsi_soa(gallivm, tokens, &params,

975

outputs);

976

else

977

lp_build_nir_soa(gallivm, shader->base.ir.nir, &params,

978

outputs);

979

980

/* Alpha test */

981

if (key->alpha.enabled) {

982

int color0 = find_output_by_semantic(&shader->info.base,

983

TGSI_SEMANTIC_COLOR,

984

0);

985

986

if (color0 != -1 && outputs[color0][3]) {

987

const struct util_format_description *cbuf_format_desc;

988

LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");

989

LLVMValueRef alpha_ref_value;

990

991

alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);

992

alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);

993

994

cbuf_format_desc = util_format_description(key->cbuf_format[0]);

995

996

lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,

997

&mask, alpha, alpha_ref_value,

998

((depth_mode & LATE_DEPTH_TEST) != 0) && !key->multisample);

999

}

1000

}

1001

1002

/* Emulate Alpha to Coverage with Alpha test */

1003

if (key->blend.alpha_to_coverage) {

1004

int color0 = find_output_by_semantic(&shader->info.base,

1005

TGSI_SEMANTIC_COLOR,

1006

0);

1007

1008

if (color0 != -1 && outputs[color0][3]) {

1009

LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");

1010

1011

if (!key->multisample) {

1012

lp_build_alpha_to_coverage(gallivm, type,

1013

&mask, alpha,

1014

(depth_mode & LATE_DEPTH_TEST) != 0);

1015

} else {

1016

lp_build_sample_alpha_to_coverage(gallivm, type, key->coverage_samples, num_loop,

1017

loop_state.counter,

1018

mask_store, alpha);

1019

}

1020

}

1021

}

1022

if (key->blend.alpha_to_one && key->multisample) {

1023

for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib) {

1024

unsigned cbuf = shader->info.base.output_semantic_index[attrib];

1025

if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&

1026

((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))

1027

if (outputs[cbuf][3]) {

1028

LLVMBuildStore(builder, lp_build_const_vec(gallivm, type, 1.0), outputs[cbuf][3]);

1029

}

1030

}

1031

}

1032

if (shader->info.base.writes_samplemask) {

1033

LLVMValueRef output_smask = NULL;

1034

int smaski = find_output_by_semantic(&shader->info.base,

1035

TGSI_SEMANTIC_SAMPLEMASK,

1036

0);

1037

struct lp_build_context smask_bld;

1038

lp_build_context_init(&smask_bld, gallivm, int_type);

1039

1040

assert(smaski >= 0);

1041

output_smask = LLVMBuildLoad(builder, outputs[smaski][0], "smask");

1042

output_smask = LLVMBuildBitCast(builder, output_smask, smask_bld.vec_type, "");

1043

if (!key->multisample && key->no_ms_sample_mask_out) {

1044

output_smask = lp_build_and(&smask_bld, output_smask, smask_bld.one);

1045

output_smask = lp_build_cmp(&smask_bld, PIPE_FUNC_NOTEQUAL, output_smask, smask_bld.zero);

1046

lp_build_mask_update(&mask, output_smask);

1047

}

1048

1049

if (key->min_samples > 1) {

1050

/* only the bit corresponding to this sample is to be used. */

1051

LLVMValueRef tmp_mask = LLVMBuildLoad(builder, out_sample_mask_storage, "tmp_mask");

1052

LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");

1053

LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, lp_build_broadcast(gallivm, int_vec_type, out_smask_idx), "");

1054

output_smask = LLVMBuildOr(builder, tmp_mask, smask_bit, "");

1055

}

1056

1057

LLVMBuildStore(builder, output_smask, out_sample_mask_storage);

1058

}

1059

1060

if (shader->info.base.writes_z) {

1061

int pos0 = find_output_by_semantic(&shader->info.base,

1062

TGSI_SEMANTIC_POSITION,

1063

0);

1064

LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos0][2], "");

1065

LLVMValueRef idx = loop_state.counter;

1066

if (key->min_samples > 1)

1067

idx = LLVMBuildAdd(builder, idx,

1068

LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");

1069

LLVMValueRef ptr = LLVMBuildGEP(builder, z_out, &idx, 1, "");

1070

LLVMBuildStore(builder, out, ptr);

1071

}

1072

1073

if (shader->info.base.writes_stencil) {

1074

int sten_out = find_output_by_semantic(&shader->info.base,

1075

TGSI_SEMANTIC_STENCIL,

1076

0);

1077

LLVMValueRef out = LLVMBuildLoad(builder, outputs[sten_out][1], "output.s");

1078

LLVMValueRef idx = loop_state.counter;

1079

if (key->min_samples > 1)

1080

idx = LLVMBuildAdd(builder, idx,

1081

LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");

1082

LLVMValueRef ptr = LLVMBuildGEP(builder, s_out, &idx, 1, "");

1083

LLVMBuildStore(builder, out, ptr);

1084

}

1085

1086

bool has_cbuf0_write = false;

1087

/* Color write - per fragment sample */

1088

for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)

1089

{

1090

unsigned cbuf = shader->info.base.output_semantic_index[attrib];

1091

if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&

1092

((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))

1093

{

1094

if (cbuf == 0 && shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS]) {

1095

/* XXX: there is an edge case with FB fetch where gl_FragColor and gl_LastFragData[0]

1096

* are used together. This creates both FRAG_RESULT_COLOR and FRAG_RESULT_DATA* output

1097

* variables. This loop then writes to cbuf 0 twice, owerwriting the correct value

1098

* from gl_FragColor with some garbage. This case is excercised in one of deqp tests.

1099

* A similar bug can happen if gl_SecondaryFragColorEXT and gl_LastFragData[1]

1100

* are mixed in the same fashion...

1101

* This workaround will break if gl_LastFragData[0] goes in outputs list before

1102

* gl_FragColor. This doesn't seem to happen though.

1103

1104

if (has_cbuf0_write)

1105

continue;

1106

has_cbuf0_write = true;

1107

}

1108

1109

for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {

1110

if(outputs[attrib][chan]) {

1111

/* XXX: just initialize outputs to point at colors[] and

1112

* skip this.

1113

1114

LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");

1115

LLVMValueRef color_ptr;

1116

LLVMValueRef color_idx = loop_state.counter;

1117

if (key->min_samples > 1)

1118

color_idx = LLVMBuildAdd(builder, color_idx,

1119

LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");

1120

color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],

1121

&color_idx, 1, "");

1122

lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);

1123

LLVMBuildStore(builder, out, color_ptr);

1124

}

1125

}

1126

}

1127

}

1128

1129

if (key->multisample && key->min_samples > 1) {

1130

LLVMBuildStore(builder, lp_build_mask_value(&mask), s_mask_ptr);

1131

lp_build_for_loop_end(&sample_loop_state);

1132

}

1133

1134

if (key->multisample) {

1135

/* execute depth test for each sample */

1136

lp_build_for_loop_begin(&sample_loop_state, gallivm,

1137

lp_build_const_int32(gallivm, 0),

1138

LLVMIntULT, lp_build_const_int32(gallivm, key->coverage_samples),

1139

lp_build_const_int32(gallivm, 1));

1140

1141

/* load the per-sample coverage mask */

1142

LLVMValueRef s_mask_idx = LLVMBuildMul(builder, sample_loop_state.counter, num_loop, "");

1143

s_mask_idx = LLVMBuildAdd(builder, s_mask_idx, loop_state.counter, "");

1144

s_mask_ptr = LLVMBuildGEP(builder, mask_store, &s_mask_idx, 1, "");

1145

1146

/* combine the execution mask post fragment shader with the coverage mask. */

1147

s_mask = LLVMBuildLoad(builder, s_mask_ptr, "");

1148

if (key->min_samples == 1)

1149

s_mask = LLVMBuildAnd(builder, s_mask, lp_build_mask_value(&mask), "");

1150

1151

/* if the shader writes sample mask use that,

1152

* but only if this isn't genuine early-depth to avoid breaking occlusion query */

1153

if (shader->info.base.writes_samplemask &&

1154

(!(depth_mode & EARLY_DEPTH_TEST) || (depth_mode & (EARLY_DEPTH_TEST_INFERRED)))) {

1155

LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");

1156

out_smask_idx = lp_build_broadcast(gallivm, int_vec_type, out_smask_idx);

1157

LLVMValueRef output_smask = LLVMBuildLoad(builder, out_sample_mask_storage, "");

1158

LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, out_smask_idx, "");

1159

LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int_vec(gallivm, int_type, 0), "");

1160

smask_bit = LLVMBuildSExt(builder, cmp, int_vec_type, "");

1161

1162

s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");

1163

}

1164

}

1165

1166

depth_ptr = depth_base_ptr;

1167

if (key->multisample) {

1168

LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_loop_state.counter, depth_sample_stride, "");

1169

depth_ptr = LLVMBuildGEP(builder, depth_ptr, &sample_offset, 1, "");

1170

}

1171

1172

/* Late Z test */

1173

if (depth_mode & LATE_DEPTH_TEST) {

1174

if (shader->info.base.writes_z) {

1175

LLVMValueRef idx = loop_state.counter;

1176

if (key->min_samples > 1)

1177

idx = LLVMBuildAdd(builder, idx,

1178

LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");

1179

LLVMValueRef ptr = LLVMBuildGEP(builder, z_out, &idx, 1, "");

1180

z = LLVMBuildLoad(builder, ptr, "output.z");

1181

} else {

1182

if (key->multisample) {

1183

lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter, key->multisample ? sample_loop_state.counter : NULL);

1184

z = interp->pos[2];

1185

}

1186

}

1187

1188

1189

* Clamp according to ARB_depth_clamp semantics.

1190

1191

if (key->depth_clamp) {

1192

z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,

1193

thread_data_ptr, z);

1194

} else {

1195

struct lp_build_context f32_bld;

1196

lp_build_context_init(&f32_bld, gallivm, type);

1197

z = lp_build_clamp(&f32_bld, z,

1198

lp_build_const_vec(gallivm, type, 0.0),

1199

lp_build_const_vec(gallivm, type, 1.0));

1200

}

1201

1202

if (shader->info.base.writes_stencil) {

1203

LLVMValueRef idx = loop_state.counter;

1204

if (key->min_samples > 1)

1205

idx = LLVMBuildAdd(builder, idx,

1206

LLVMBuildMul(builder, sample_loop_state.counter, num_loop, ""), "");

1207

LLVMValueRef ptr = LLVMBuildGEP(builder, s_out, &idx, 1, "");

1208

stencil_refs[0] = LLVMBuildLoad(builder, ptr, "output.s");

1209

/* there's only one value, and spec says to discard additional bits */

1210

LLVMValueRef s_max_mask = lp_build_const_int_vec(gallivm, int_type, 255);

1211

stencil_refs[0] = LLVMBuildBitCast(builder, stencil_refs[0], int_vec_type, "");

1212

stencil_refs[0] = LLVMBuildAnd(builder, stencil_refs[0], s_max_mask, "");

1213

stencil_refs[1] = stencil_refs[0];

1214

}

1215

1216

lp_build_depth_stencil_load_swizzled(gallivm, type,

1217

zs_format_desc, key->resource_1d,

1218

depth_ptr, depth_stride,

1219

&z_fb, &s_fb, loop_state.counter);

1220

1221

lp_build_depth_stencil_test(gallivm,

1222

&key->depth,

1223

key->stencil,

1224

type,

1225

zs_format_desc,

1226

key->multisample ? NULL : &mask,

1227

&s_mask,

1228

stencil_refs,

1229

z, z_fb, s_fb,

1230

facing,

1231

&z_value, &s_value,

1232

!simple_shader);

1233

/* Late Z write */

1234

if (depth_mode & LATE_DEPTH_WRITE) {

1235

lp_build_depth_stencil_write_swizzled(gallivm, type,

1236

zs_format_desc, key->resource_1d,

1237

NULL, NULL, NULL, loop_state.counter,

1238

depth_ptr, depth_stride,

1239

z_value, s_value);

1240

}

1241

}

1242

else if ((depth_mode & EARLY_DEPTH_TEST) &&

1243

(depth_mode & LATE_DEPTH_WRITE))

1244

{

1245

/* Need to apply a reduced mask to the depth write. Reload the

1246

* depth value, update from zs_value with the new mask value and

1247

* write that out.

1248

1249

if (key->multisample) {

1250

z_value = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_sample_value_store, sample_loop_state.counter), z_type, "");;

1251

s_value = lp_build_pointer_get(builder, s_sample_value_store, sample_loop_state.counter);

1252

z_fb = LLVMBuildBitCast(builder, lp_build_pointer_get(builder, z_fb_store, sample_loop_state.counter), z_fb_type, "");

1253

s_fb = lp_build_pointer_get(builder, s_fb_store, sample_loop_state.counter);

1254

}

1255

lp_build_depth_stencil_write_swizzled(gallivm, type,

1256

zs_format_desc, key->resource_1d,

1257

key->multisample ? s_mask : lp_build_mask_value(&mask), z_fb, s_fb, loop_state.counter,

1258

depth_ptr, depth_stride,

1259

z_value, s_value);

1260

}

1261

1262

if (key->occlusion_count) {

1263

LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);

1264

lp_build_name(counter, "counter");

1265

1266

lp_build_occlusion_count(gallivm, type,

1267

key->multisample ? s_mask : lp_build_mask_value(&mask), counter);

1268

}

1269

1270

/* if this is genuine early-depth in the shader, write samplemask now

1271

* after occlusion count has been updated

1272

1273

if (key->multisample && shader->info.base.writes_samplemask &&

1274

(depth_mode & (EARLY_DEPTH_TEST_INFERRED | EARLY_DEPTH_TEST)) == EARLY_DEPTH_TEST) {

1275

/* if the shader writes sample mask use that */

1276

LLVMValueRef out_smask_idx = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1), sample_loop_state.counter, "");

1277

out_smask_idx = lp_build_broadcast(gallivm, int_vec_type, out_smask_idx);

1278

LLVMValueRef output_smask = LLVMBuildLoad(builder, out_sample_mask_storage, "");

1279

LLVMValueRef smask_bit = LLVMBuildAnd(builder, output_smask, out_smask_idx, "");

1280

LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int_vec(gallivm, int_type, 0), "");

1281

smask_bit = LLVMBuildSExt(builder, cmp, int_vec_type, "");

1282

1283

s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");

1284

}

1285

1286

1287

if (key->multisample) {

1288

/* store the sample mask for this loop */

1289

LLVMBuildStore(builder, s_mask, s_mask_ptr);

1290

lp_build_for_loop_end(&sample_loop_state);

1291

}

1292

1293

mask_val = lp_build_mask_end(&mask);

1294

if (!key->multisample)

1295

LLVMBuildStore(builder, mask_val, mask_ptr);

1296

lp_build_for_loop_end(&loop_state);

1297

}

1298

1299

1300

/**

1301

* This function will reorder pixels from the fragment shader SoA to memory layout AoS

1302

1303

* Fragment Shader outputs pixels in small 2x2 blocks

1304

* e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...

1305

1306

* However in memory pixels are stored in rows

1307

* e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...

1308

1309

* @param type fragment shader type (4x or 8x float)

1310

* @param num_fs number of fs_src

1311

* @param is_1d whether we're outputting to a 1d resource

1312

* @param dst_channels number of output channels

1313

* @param fs_src output from fragment shader

1314

* @param dst pointer to store result

1315

* @param pad_inline is channel padding inline or at end of row

1316

* @return the number of dsts

1317

1318

static int

1319

generate_fs_twiddle(struct gallivm_state *gallivm,

1320

struct lp_type type,

1321

unsigned num_fs,

1322

unsigned dst_channels,

1323

LLVMValueRef fs_src[][4],

1324

LLVMValueRef* dst,

1325

bool pad_inline)

1326

{

1327

LLVMValueRef src[16];

1328

1329

bool swizzle_pad;

1330

bool twiddle;

1331

bool split;

1332

1333

unsigned pixels = type.length / 4;

1334

unsigned reorder_group;

1335

unsigned src_channels;

1336

unsigned src_count;

1337

unsigned i;

1338

1339

src_channels = dst_channels < 3 ? dst_channels : 4;

1340

src_count = num_fs * src_channels;

1341

1342

assert(pixels == 2 || pixels == 1);

1343

assert(num_fs * src_channels <= ARRAY_SIZE(src));

1344

1345

1346

* Transpose from SoA -> AoS

1347

1348

for (i = 0; i < num_fs; ++i) {

1349

lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);

1350

}

1351

1352

1353

* Pick transformation options

1354

1355

swizzle_pad = false;

1356

twiddle = false;

1357

split = false;

1358

reorder_group = 0;

1359

1360

if (dst_channels == 1) {

1361

twiddle = true;

1362

1363

if (pixels == 2) {

1364

split = true;

1365

}

1366

} else if (dst_channels == 2) {

1367

if (pixels == 1) {

1368

reorder_group = 1;

1369

}

1370

} else if (dst_channels > 2) {

1371

if (pixels == 1) {

1372

reorder_group = 2;

1373

} else {

1374

twiddle = true;

1375

}

1376

1377

if (!pad_inline && dst_channels == 3 && pixels > 1) {

1378

swizzle_pad = true;

1379

}

1380

}

1381

1382

1383

* Split the src in half

1384

1385

if (split) {

1386

for (i = num_fs; i > 0; --i) {

1387

src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);

1388

src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);

1389

}

1390

1391

src_count *= 2;

1392

type.length = 4;

1393

}

1394

1395

1396

* Ensure pixels are in memory order

1397

1398

if (reorder_group) {

1399

/* Twiddle pixels by reordering the array, e.g.:

1400

1401

* src_count = 8 -> 0 2 1 3 4 6 5 7

1402

* src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15

1403

1404

const unsigned reorder_sw[] = { 0, 2, 1, 3 };

1405

1406

for (i = 0; i < src_count; ++i) {

1407

unsigned group = i / reorder_group;

1408

unsigned block = (group / 4) * 4 * reorder_group;

1409

unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);

1410

dst[i] = src[j];

1411

}

1412

} else if (twiddle) {

1413

/* Twiddle pixels across elements of array */

1414

1415

* XXX: we should avoid this in some cases, but would need to tell

1416

* lp_build_conv to reorder (or deal with it ourselves).

1417

1418

lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);

1419

} else {

1420

/* Do nothing */

1421

memcpy(dst, src, sizeof(LLVMValueRef) * src_count);

1422

}

1423

1424

1425

* Moves any padding between pixels to the end

1426

* e.g. RGBXRGBX -> RGBRGBXX

1427

1428

if (swizzle_pad) {

1429

unsigned char swizzles[16];

1430

unsigned elems = pixels * dst_channels;

1431

1432

for (i = 0; i < type.length; ++i) {

1433

if (i < elems)

1434

swizzles[i] = i % dst_channels + (i / dst_channels) * 4;

1435

else

1436

swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;

1437

}

1438

1439

for (i = 0; i < src_count; ++i) {

1440

dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);

1441

}

1442

}

1443

1444

return src_count;

1445

}

1446

1447

1448

1449

* Untwiddle and transpose, much like the above.

1450

* However, this is after conversion, so we get packed vectors.

1451

* At this time only handle 4x16i8 rgba / 2x16i8 rg / 1x16i8 r data,

1452

* the vectors will look like:

1453

* r0r1r4r5r2r3r6r7r8r9r12... (albeit color channels may

1454

* be swizzled here). Extending to 16bit should be trivial.

1455

* Should also be extended to handle twice wide vectors with AVX2...

1456

1457

static void

1458

fs_twiddle_transpose(struct gallivm_state *gallivm,

1459

struct lp_type type,

1460

LLVMValueRef *src,

1461

unsigned src_count,

1462

LLVMValueRef *dst)

1463

{

1464

unsigned i, j;

1465

struct lp_type type64, type16, type32;

1466

LLVMTypeRef type64_t, type8_t, type16_t, type32_t;

1467

LLVMBuilderRef builder = gallivm->builder;

1468

LLVMValueRef tmp[4], shuf[8];

1469

for (j = 0; j < 2; j++) {

1470

shuf[j*4 + 0] = lp_build_const_int32(gallivm, j*4 + 0);

1471

shuf[j*4 + 1] = lp_build_const_int32(gallivm, j*4 + 2);

1472

shuf[j*4 + 2] = lp_build_const_int32(gallivm, j*4 + 1);

1473

shuf[j*4 + 3] = lp_build_const_int32(gallivm, j*4 + 3);

1474

}

1475

1476

assert(src_count == 4 || src_count == 2 || src_count == 1);

1477

assert(type.width == 8);

1478

assert(type.length == 16);

1479

1480

type8_t = lp_build_vec_type(gallivm, type);

1481

1482

type64 = type;

1483

type64.length /= 8;

1484

type64.width *= 8;

1485

type64_t = lp_build_vec_type(gallivm, type64);

1486

1487

type16 = type;

1488

type16.length /= 2;

1489

type16.width *= 2;

1490

type16_t = lp_build_vec_type(gallivm, type16);

1491

1492

type32 = type;

1493

type32.length /= 4;

1494

type32.width *= 4;

1495

type32_t = lp_build_vec_type(gallivm, type32);

1496

1497

lp_build_transpose_aos_n(gallivm, type, src, src_count, tmp);

1498

1499

if (src_count == 1) {

1500

/* transpose was no-op, just untwiddle */

1501

LLVMValueRef shuf_vec;

1502

shuf_vec = LLVMConstVector(shuf, 8);

1503

tmp[0] = LLVMBuildBitCast(builder, src[0], type16_t, "");

1504

tmp[0] = LLVMBuildShuffleVector(builder, tmp[0], tmp[0], shuf_vec, "");

1505

dst[0] = LLVMBuildBitCast(builder, tmp[0], type8_t, "");

1506

} else if (src_count == 2) {

1507

LLVMValueRef shuf_vec;

1508

shuf_vec = LLVMConstVector(shuf, 4);

1509

1510

for (i = 0; i < 2; i++) {

1511

tmp[i] = LLVMBuildBitCast(builder, tmp[i], type32_t, "");

1512

tmp[i] = LLVMBuildShuffleVector(builder, tmp[i], tmp[i], shuf_vec, "");

1513

dst[i] = LLVMBuildBitCast(builder, tmp[i], type8_t, "");

1514

}

1515

} else {

1516

for (j = 0; j < 2; j++) {

1517

LLVMValueRef lo, hi, lo2, hi2;

1518

1519

* Note that if we only really have 3 valid channels (rgb)

1520

* and we don't need alpha we could substitute a undef here

1521

* for the respective channel (causing llvm to drop conversion

1522

* for alpha).

1523

1524

/* we now have rgba0rgba1rgba4rgba5 etc, untwiddle */

1525

lo2 = LLVMBuildBitCast(builder, tmp[j*2], type64_t, "");

1526

hi2 = LLVMBuildBitCast(builder, tmp[j*2 + 1], type64_t, "");

1527

lo = lp_build_interleave2(gallivm, type64, lo2, hi2, 0);

1528

hi = lp_build_interleave2(gallivm, type64, lo2, hi2, 1);

1529

dst[j*2] = LLVMBuildBitCast(builder, lo, type8_t, "");

1530

dst[j*2 + 1] = LLVMBuildBitCast(builder, hi, type8_t, "");

1531

}

1532

}

1533

}

1534

1535

1536

/**

1537

* Load an unswizzled block of pixels from memory

1538

1539

static void

1540

load_unswizzled_block(struct gallivm_state *gallivm,

1541

LLVMValueRef base_ptr,

1542

LLVMValueRef stride,

1543

unsigned block_width,

1544

unsigned block_height,

1545

LLVMValueRef* dst,

1546

struct lp_type dst_type,

1547

unsigned dst_count,

1548

unsigned dst_alignment)

1549

{

1550

LLVMBuilderRef builder = gallivm->builder;

1551

unsigned row_size = dst_count / block_height;

1552

unsigned i;

1553

1554

/* Ensure block exactly fits into dst */

1555

assert((block_width * block_height) % dst_count == 0);

1556

1557

for (i = 0; i < dst_count; ++i) {

1558

unsigned x = i % row_size;

1559

unsigned y = i / row_size;

1560

1561

LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);

1562

LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");

1563

1564

LLVMValueRef gep[2];

1565

LLVMValueRef dst_ptr;

1566

1567

gep[0] = lp_build_const_int32(gallivm, 0);

1568

gep[1] = LLVMBuildAdd(builder, bx, by, "");

1569

1570

dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");

1571

dst_ptr = LLVMBuildBitCast(builder, dst_ptr,

1572

LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");

1573

1574

dst[i] = LLVMBuildLoad(builder, dst_ptr, "");

1575

1576

LLVMSetAlignment(dst[i], dst_alignment);

1577

}

1578

}

1579

1580

1581

/**

1582

* Store an unswizzled block of pixels to memory

1583

1584

static void

1585

store_unswizzled_block(struct gallivm_state *gallivm,

1586

LLVMValueRef base_ptr,

1587

LLVMValueRef stride,

1588

unsigned block_width,

1589

unsigned block_height,

1590

LLVMValueRef* src,

1591

struct lp_type src_type,

1592

unsigned src_count,

1593

unsigned src_alignment)

1594

{

1595

LLVMBuilderRef builder = gallivm->builder;

1596

unsigned row_size = src_count / block_height;

1597

unsigned i;

1598

1599

/* Ensure src exactly fits into block */

1600

assert((block_width * block_height) % src_count == 0);

1601

1602

for (i = 0; i < src_count; ++i) {

1603

unsigned x = i % row_size;

1604

unsigned y = i / row_size;

1605

1606

LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);

1607

LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");

1608

1609

LLVMValueRef gep[2];

1610

LLVMValueRef src_ptr;

1611

1612

gep[0] = lp_build_const_int32(gallivm, 0);

1613

gep[1] = LLVMBuildAdd(builder, bx, by, "");

1614

1615

src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");

1616

src_ptr = LLVMBuildBitCast(builder, src_ptr,

1617

LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");

1618

1619

src_ptr = LLVMBuildStore(builder, src[i], src_ptr);

1620

1621

LLVMSetAlignment(src_ptr, src_alignment);

1622

}

1623

}

1624

1625

1626

1627

/**

1628

* Retrieves the type for a format which is usable in the blending code.

1629

1630

* e.g. RGBA16F = 4x float, R3G3B2 = 3x byte

1631

1632

static inline void

1633

lp_blend_type_from_format_desc(const struct util_format_description *format_desc,

1634

struct lp_type* type)

1635

{

1636

unsigned i;

1637

unsigned chan;

1638

1639

if (format_expands_to_float_soa(format_desc)) {

1640

/* always use ordinary floats for blending */

1641

type->floating = true;

1642

type->fixed = false;

1643

type->sign = true;

1644

type->norm = false;

1645

type->width = 32;

1646

type->length = 4;

1647

return;

1648

}

1649

1650

for (i = 0; i < 4; i++)

1651

if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)

1652

break;

1653

chan = i;

1654

1655

memset(type, 0, sizeof(struct lp_type));

1656

type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;

1657

type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;

1658

type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;

1659

type->norm = format_desc->channel[chan].normalized;

1660

type->width = format_desc->channel[chan].size;

1661

type->length = format_desc->nr_channels;

1662

1663

for (i = 1; i < format_desc->nr_channels; ++i) {

1664

if (format_desc->channel[i].size > type->width)

1665

type->width = format_desc->channel[i].size;

1666

}

1667

1668

if (type->floating) {

1669

type->width = 32;

1670

} else {

1671

if (type->width <= 8) {

1672

type->width = 8;

1673

} else if (type->width <= 16) {

1674

type->width = 16;

1675

} else {

1676

type->width = 32;

1677

}

1678

}

1679

1680

if (is_arithmetic_format(format_desc) && type->length == 3) {

1681

type->length = 4;

1682

}

1683

}

1684

1685

1686

/**

1687

* Scale a normalized value from src_bits to dst_bits.

1688

1689

* The exact calculation is

1690

1691

* dst = iround(src * dst_mask / src_mask)

1692

1693

* or with integer rounding

1694

1695

* dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)

1696

1697

* where

1698

1699

* src_mask = (1 << src_bits) - 1

1700

* dst_mask = (1 << dst_bits) - 1

1701

1702

* but we try to avoid division and multiplication through shifts.

1703

1704

static inline LLVMValueRef

1705

scale_bits(struct gallivm_state *gallivm,

1706

int src_bits,

1707

int dst_bits,

1708

LLVMValueRef src,

1709

struct lp_type src_type)

1710

{

1711

LLVMBuilderRef builder = gallivm->builder;

1712

LLVMValueRef result = src;

1713

1714

if (dst_bits < src_bits) {

1715

int delta_bits = src_bits - dst_bits;

1716

1717

if (delta_bits <= dst_bits) {

1718

1719

if (dst_bits == 4) {

1720

struct lp_type flt_type = lp_type_float_vec(32, src_type.length * 32);

1721

1722

result = lp_build_unsigned_norm_to_float(gallivm, src_bits, flt_type, src);

1723

result = lp_build_clamped_float_to_unsigned_norm(gallivm, flt_type, dst_bits, result);

1724

result = LLVMBuildTrunc(gallivm->builder, result, lp_build_int_vec_type(gallivm, src_type), "");

1725

return result;

1726

}

1727

1728

1729

* Approximate the rescaling with a single shift.

1730

1731

* This gives the wrong rounding.

1732

1733

1734

result = LLVMBuildLShr(builder,

1735

src,

1736

lp_build_const_int_vec(gallivm, src_type, delta_bits),

1737

"");

1738

1739

} else {

1740

1741

* Try more accurate rescaling.

1742

1743

1744

1745

* Drop the least significant bits to make space for the multiplication.

1746

1747

* XXX: A better approach would be to use a wider integer type as intermediate. But

1748

* this is enough to convert alpha from 16bits -> 2 when rendering to

1749

* PIPE_FORMAT_R10G10B10A2_UNORM.

1750

1751

result = LLVMBuildLShr(builder,

1752

src,

1753

lp_build_const_int_vec(gallivm, src_type, dst_bits),

1754

"");

1755

1756

1757

result = LLVMBuildMul(builder,

1758

result,

1759

lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),

1760

"");

1761

1762

1763

* Add a rounding term before the division.

1764

1765

* TODO: Handle signed integers too.

1766

1767

if (!src_type.sign) {

1768

result = LLVMBuildAdd(builder,

1769

result,

1770

lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),

1771

"");

1772

}

1773

1774

1775

* Approximate the division by src_mask with a src_bits shift.

1776

1777

* Given the src has already been shifted by dst_bits, all we need

1778

* to do is to shift by the difference.

1779

1780

1781

result = LLVMBuildLShr(builder,

1782

result,

1783

lp_build_const_int_vec(gallivm, src_type, delta_bits),

1784

"");

1785

}

1786

1787

} else if (dst_bits > src_bits) {

1788

/* Scale up bits */

1789

int db = dst_bits - src_bits;

1790

1791

/* Shift left by difference in bits */

1792

result = LLVMBuildShl(builder,

1793

src,

1794

lp_build_const_int_vec(gallivm, src_type, db),

1795

"");

1796

1797

if (db <= src_bits) {

1798

/* Enough bits in src to fill the remainder */

1799

LLVMValueRef lower = LLVMBuildLShr(builder,

1800

src,

1801

lp_build_const_int_vec(gallivm, src_type, src_bits - db),

1802

"");

1803

1804

result = LLVMBuildOr(builder, result, lower, "");

1805

} else if (db > src_bits) {

1806

/* Need to repeatedly copy src bits to fill remainder in dst */

1807

unsigned n;

1808

1809

for (n = src_bits; n < dst_bits; n *= 2) {

1810

LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);

1811

1812

result = LLVMBuildOr(builder,

1813

result,

1814

LLVMBuildLShr(builder, result, shuv, ""),

1815

"");

1816

}

1817

}

1818

}

1819

1820

return result;

1821

}

1822

1823

/**

1824

* If RT is a smallfloat (needing denorms) format

1825

1826

static inline int

1827

have_smallfloat_format(struct lp_type dst_type,

1828

enum pipe_format format)

1829

{

1830

return ((dst_type.floating && dst_type.width != 32) ||

1831

/* due to format handling hacks this format doesn't have floating set

1832

* here (and actually has width set to 32 too) so special case this. */

1833

(format == PIPE_FORMAT_R11G11B10_FLOAT));

1834

}

1835

1836

1837

/**

1838

* Convert from memory format to blending format

1839

1840

* e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending

1841

1842

static void

1843

convert_to_blend_type(struct gallivm_state *gallivm,

1844

unsigned block_size,

1845

const struct util_format_description *src_fmt,

1846

struct lp_type src_type,

1847

struct lp_type dst_type,

1848

LLVMValueRef* src, // and dst

1849

unsigned num_srcs)

1850

{

1851

LLVMValueRef *dst = src;

1852

LLVMBuilderRef builder = gallivm->builder;

1853

struct lp_type blend_type;

1854

struct lp_type mem_type;

1855

unsigned i, j;

1856

unsigned pixels = block_size / num_srcs;

1857

bool is_arith;

1858

1859

1860

* full custom path for packed floats and srgb formats - none of the later

1861

* functions would do anything useful, and given the lp_type representation they

1862

* can't be fixed. Should really have some SoA blend path for these kind of

1863

* formats rather than hacking them in here.

1864

1865

if (format_expands_to_float_soa(src_fmt)) {

1866

LLVMValueRef tmpsrc[4];

1867

1868

* This is pretty suboptimal for this case blending in SoA would be much

1869

* better, since conversion gets us SoA values so need to convert back.

1870

1871

assert(src_type.width == 32 || src_type.width == 16);

1872

assert(dst_type.floating);

1873

assert(dst_type.width == 32);

1874

assert(dst_type.length % 4 == 0);

1875

assert(num_srcs % 4 == 0);

1876

1877

if (src_type.width == 16) {

1878

/* expand 4x16bit values to 4x32bit */

1879

struct lp_type type32x4 = src_type;

1880

LLVMTypeRef ltype32x4;

1881

unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;

1882

type32x4.width = 32;

1883

ltype32x4 = lp_build_vec_type(gallivm, type32x4);

1884

for (i = 0; i < num_fetch; i++) {

1885

src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");

1886

}

1887

src_type.width = 32;

1888

}

1889

for (i = 0; i < 4; i++) {

1890

tmpsrc[i] = src[i];

1891

}

1892

for (i = 0; i < num_srcs / 4; i++) {

1893

LLVMValueRef tmpsoa[4];

1894

LLVMValueRef tmps = tmpsrc[i];

1895

if (dst_type.length == 8) {

1896

LLVMValueRef shuffles[8];

1897

unsigned j;

1898

/* fetch was 4 values but need 8-wide output values */

1899

tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);

1900

1901

* for 8-wide aos transpose would give us wrong order not matching

1902

* incoming converted fs values and mask. ARGH.

1903

1904

for (j = 0; j < 4; j++) {

1905

shuffles[j] = lp_build_const_int32(gallivm, j * 2);

1906

shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);

1907

}

1908

tmps = LLVMBuildShuffleVector(builder, tmps, tmps,

1909

LLVMConstVector(shuffles, 8), "");

1910

}

1911

if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {

1912

lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);

1913

}

1914

else {

1915

lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);

1916

}

1917

lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);

1918

}

1919

return;

1920

}

1921

1922

lp_mem_type_from_format_desc(src_fmt, &mem_type);

1923

lp_blend_type_from_format_desc(src_fmt, &blend_type);

1924

1925

/* Is the format arithmetic */

1926

is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;

1927

is_arith &= !(mem_type.width == 16 && mem_type.floating);

1928

1929

/* Pad if necessary */

1930

if (!is_arith && src_type.length < dst_type.length) {

1931

for (i = 0; i < num_srcs; ++i) {

1932

dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);

1933

}

1934

1935

src_type.length = dst_type.length;

1936

}

1937

1938

/* Special case for half-floats */

1939

if (mem_type.width == 16 && mem_type.floating) {

1940

assert(blend_type.width == 32 && blend_type.floating);

1941

lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);

1942

is_arith = false;

1943

}

1944

1945

if (!is_arith) {

1946

return;

1947

}

1948

1949

src_type.width = blend_type.width * blend_type.length;

1950

blend_type.length *= pixels;

1951

src_type.length *= pixels / (src_type.length / mem_type.length);

1952

1953

for (i = 0; i < num_srcs; ++i) {

1954

LLVMValueRef chans;

1955

LLVMValueRef res = NULL;

1956

1957

dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");

1958

1959

for (j = 0; j < src_fmt->nr_channels; ++j) {

1960

unsigned mask = 0;

1961

unsigned sa = src_fmt->channel[j].shift;

1962

#if UTIL_ARCH_LITTLE_ENDIAN

1963

unsigned from_lsb = j;

1964

#else

1965

unsigned from_lsb = (blend_type.length / pixels) - j - 1;

1966

#endif

1967

1968

mask = (1 << src_fmt->channel[j].size) - 1;

1969

1970

/* Extract bits from source */

1971

chans = LLVMBuildLShr(builder,

1972

dst[i],

1973

lp_build_const_int_vec(gallivm, src_type, sa),

1974

"");

1975

1976

chans = LLVMBuildAnd(builder,

1977

chans,

1978

lp_build_const_int_vec(gallivm, src_type, mask),

1979

"");

1980

1981

/* Scale bits */

1982

if (src_type.norm) {

1983

chans = scale_bits(gallivm, src_fmt->channel[j].size,

1984

blend_type.width, chans, src_type);

1985

}

1986

1987

/* Insert bits into correct position */

1988

chans = LLVMBuildShl(builder,

1989

chans,

1990

lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),

1991

"");

1992

1993

if (j == 0) {

1994

res = chans;

1995

} else {

1996

res = LLVMBuildOr(builder, res, chans, "");

1997

}

1998

}

1999

2000

dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");

2001

}

2002

}

2003

2004

2005

/**

2006

* Convert from blending format to memory format

2007

2008

* e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory

2009

2010

static void

2011

convert_from_blend_type(struct gallivm_state *gallivm,

2012

unsigned block_size,

2013

const struct util_format_description *src_fmt,

2014

struct lp_type src_type,

2015

struct lp_type dst_type,

2016

LLVMValueRef* src, // and dst

2017

unsigned num_srcs)

2018

{

2019

LLVMValueRef* dst = src;

2020

unsigned i, j, k;

2021

struct lp_type mem_type;

2022

struct lp_type blend_type;

2023

LLVMBuilderRef builder = gallivm->builder;

2024

unsigned pixels = block_size / num_srcs;

2025

bool is_arith;

2026

2027

2028

* full custom path for packed floats and srgb formats - none of the later

2029

* functions would do anything useful, and given the lp_type representation they

2030

* can't be fixed. Should really have some SoA blend path for these kind of

2031

* formats rather than hacking them in here.

2032

2033

if (format_expands_to_float_soa(src_fmt)) {

2034

2035

* This is pretty suboptimal for this case blending in SoA would be much

2036

* better - we need to transpose the AoS values back to SoA values for

2037

* conversion/packing.

2038

2039

assert(src_type.floating);

2040

assert(src_type.width == 32);

2041

assert(src_type.length % 4 == 0);

2042

assert(dst_type.width == 32 || dst_type.width == 16);

2043

2044

for (i = 0; i < num_srcs / 4; i++) {

2045

LLVMValueRef tmpsoa[4], tmpdst;

2046

lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);

2047

/* really really need SoA here */

2048

2049

if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {

2050

tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);

2051

}

2052

else {

2053

tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,

2054

src_type, tmpsoa);

2055

}

2056

2057

if (src_type.length == 8) {

2058

LLVMValueRef tmpaos, shuffles[8];

2059

unsigned j;

2060

2061

* for 8-wide aos transpose has given us wrong order not matching

2062

* output order. HMPF. Also need to split the output values manually.

2063

2064

for (j = 0; j < 4; j++) {

2065

shuffles[j * 2] = lp_build_const_int32(gallivm, j);

2066

shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);

2067

}

2068

tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,

2069

LLVMConstVector(shuffles, 8), "");

2070

src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);

2071

src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);

2072

}

2073

else {

2074

src[i] = tmpdst;

2075

}

2076

}

2077

if (dst_type.width == 16) {

2078

struct lp_type type16x8 = dst_type;

2079

struct lp_type type32x4 = dst_type;

2080

LLVMTypeRef ltype16x4, ltypei64, ltypei128;

2081

unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;

2082

type16x8.length = 8;

2083

type32x4.width = 32;

2084

ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);

2085

ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);

2086

ltype16x4 = lp_build_vec_type(gallivm, dst_type);

2087

/* We could do vector truncation but it doesn't generate very good code */

2088

for (i = 0; i < num_fetch; i++) {

2089

src[i] = lp_build_pack2(gallivm, type32x4, type16x8,

2090

src[i], lp_build_zero(gallivm, type32x4));

2091

src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");

2092

src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");

2093

src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");

2094

}

2095

}

2096

return;

2097

}

2098

2099

lp_mem_type_from_format_desc(src_fmt, &mem_type);

2100

lp_blend_type_from_format_desc(src_fmt, &blend_type);

2101

2102

is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);

2103

2104

/* Special case for half-floats */

2105

if (mem_type.width == 16 && mem_type.floating) {

2106

int length = dst_type.length;

2107

assert(blend_type.width == 32 && blend_type.floating);

2108

2109

dst_type.length = src_type.length;

2110

2111

lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);

2112

2113

dst_type.length = length;

2114

is_arith = false;

2115

}

2116

2117

/* Remove any padding */

2118

if (!is_arith && (src_type.length % mem_type.length)) {

2119

src_type.length -= (src_type.length % mem_type.length);

2120

2121

for (i = 0; i < num_srcs; ++i) {

2122

dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);

2123

}

2124

}

2125

2126

/* No bit arithmetic to do */

2127

if (!is_arith) {

2128

return;

2129

}

2130

2131

src_type.length = pixels;

2132

src_type.width = blend_type.length * blend_type.width;

2133

dst_type.length = pixels;

2134

2135

for (i = 0; i < num_srcs; ++i) {

2136

LLVMValueRef chans;

2137

LLVMValueRef res = NULL;

2138

2139

dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");

2140

2141

for (j = 0; j < src_fmt->nr_channels; ++j) {

2142

unsigned mask = 0;

2143

unsigned sa = src_fmt->channel[j].shift;

2144

unsigned sz_a = src_fmt->channel[j].size;

2145

#if UTIL_ARCH_LITTLE_ENDIAN

2146

unsigned from_lsb = j;

2147

#else

2148

unsigned from_lsb = blend_type.length - j - 1;

2149

#endif

2150

2151

assert(blend_type.width > src_fmt->channel[j].size);

2152

2153

for (k = 0; k < blend_type.width; ++k) {

2154

mask |= 1 << k;

2155

}

2156

2157

/* Extract bits */

2158

chans = LLVMBuildLShr(builder,

2159

dst[i],

2160

lp_build_const_int_vec(gallivm, src_type,

2161

from_lsb * blend_type.width),

2162

"");

2163

2164

chans = LLVMBuildAnd(builder,

2165

chans,

2166

lp_build_const_int_vec(gallivm, src_type, mask),

2167

"");

2168

2169

/* Scale down bits */

2170

if (src_type.norm) {

2171

chans = scale_bits(gallivm, blend_type.width,

2172

src_fmt->channel[j].size, chans, src_type);

2173

} else if (!src_type.floating && sz_a < blend_type.width) {

2174

LLVMValueRef mask_val = lp_build_const_int_vec(gallivm, src_type, (1UL << sz_a) - 1);

2175

LLVMValueRef mask = LLVMBuildICmp(builder, LLVMIntUGT, chans, mask_val, "");

2176

chans = LLVMBuildSelect(builder, mask, mask_val, chans, "");

2177

}

2178

2179

/* Insert bits */

2180

chans = LLVMBuildShl(builder,

2181

chans,

2182

lp_build_const_int_vec(gallivm, src_type, sa),

2183

"");

2184

2185

sa += src_fmt->channel[j].size;

2186

2187

if (j == 0) {

2188

res = chans;

2189

} else {

2190

res = LLVMBuildOr(builder, res, chans, "");

2191

}

2192

}

2193

2194

assert (dst_type.width != 24);

2195

2196

dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");

2197

}

2198

}

2199

2200

2201

/**

2202

* Convert alpha to same blend type as src

2203

2204

static void

2205

convert_alpha(struct gallivm_state *gallivm,

2206

struct lp_type row_type,

2207

struct lp_type alpha_type,

2208

const unsigned block_size,

2209

const unsigned block_height,

2210

const unsigned src_count,

2211

const unsigned dst_channels,

2212

const bool pad_inline,

2213

LLVMValueRef* src_alpha)

2214

{

2215

LLVMBuilderRef builder = gallivm->builder;

2216

unsigned i, j;

2217

unsigned length = row_type.length;

2218

row_type.length = alpha_type.length;

2219

2220

/* Twiddle the alpha to match pixels */

2221

lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);

2222

2223

2224

* TODO this should use single lp_build_conv call for

2225

* src_count == 1 && dst_channels == 1 case (dropping the concat below)

2226

2227

for (i = 0; i < block_height; ++i) {

2228

lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);

2229

}

2230

2231

alpha_type = row_type;

2232

row_type.length = length;

2233

2234

/* If only one channel we can only need the single alpha value per pixel */

2235

if (src_count == 1 && dst_channels == 1) {

2236

2237

lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);

2238

} else {

2239

/* If there are more srcs than rows then we need to split alpha up */

2240

if (src_count > block_height) {

2241

for (i = src_count; i > 0; --i) {

2242

unsigned pixels = block_size / src_count;

2243

unsigned idx = i - 1;

2244

2245

src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],

2246

(idx * pixels) % 4, pixels);

2247

}

2248

}

2249

2250

/* If there is a src for each pixel broadcast the alpha across whole row */

2251

if (src_count == block_size) {

2252

for (i = 0; i < src_count; ++i) {

2253

src_alpha[i] = lp_build_broadcast(gallivm,

2254

lp_build_vec_type(gallivm, row_type), src_alpha[i]);

2255

}

2256

} else {

2257

unsigned pixels = block_size / src_count;

2258

unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;

2259

unsigned alpha_span = 1;

2260

LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];

2261

2262

/* Check if we need 2 src_alphas for our shuffles */

2263

if (pixels > alpha_type.length) {

2264

alpha_span = 2;

2265

}

2266

2267

/* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */

2268

for (j = 0; j < row_type.length; ++j) {

2269

if (j < pixels * channels) {

2270

shuffles[j] = lp_build_const_int32(gallivm, j / channels);

2271

} else {

2272

shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));

2273

}

2274

}

2275

2276

for (i = 0; i < src_count; ++i) {

2277

unsigned idx1 = i, idx2 = i;

2278

2279

if (alpha_span > 1){

2280

idx1 *= alpha_span;

2281

idx2 = idx1 + 1;

2282

}

2283

2284

src_alpha[i] = LLVMBuildShuffleVector(builder,

2285

src_alpha[idx1],

2286

src_alpha[idx2],

2287

LLVMConstVector(shuffles, row_type.length),

2288

"");

2289

}

2290

}

2291

}

2292

}

2293

2294

2295

/**

2296

* Generates the blend function for unswizzled colour buffers

2297

* Also generates the read & write from colour buffer

2298

2299

static void

2300

generate_unswizzled_blend(struct gallivm_state *gallivm,

2301

unsigned rt,

2302

struct lp_fragment_shader_variant *variant,

2303

enum pipe_format out_format,

2304

unsigned int num_fs,

2305

struct lp_type fs_type,

2306

LLVMValueRef* fs_mask,

2307

LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],

2308

LLVMValueRef context_ptr,

2309

LLVMValueRef color_ptr,

2310

LLVMValueRef stride,

2311

unsigned partial_mask,

2312

boolean do_branch)

2313

{

2314

const unsigned alpha_channel = 3;

2315

const unsigned block_width = LP_RASTER_BLOCK_SIZE;

2316

const unsigned block_height = LP_RASTER_BLOCK_SIZE;

2317

const unsigned block_size = block_width * block_height;

2318

const unsigned lp_integer_vector_width = 128;

2319

2320

LLVMBuilderRef builder = gallivm->builder;

2321

LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];

2322

LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];

2323

LLVMValueRef src_alpha[4 * 4];

2324

LLVMValueRef src1_alpha[4 * 4] = { NULL };

2325

LLVMValueRef src_mask[4 * 4];

2326

LLVMValueRef src[4 * 4];

2327

LLVMValueRef src1[4 * 4];

2328

LLVMValueRef dst[4 * 4];

2329

LLVMValueRef blend_color;

2330

LLVMValueRef blend_alpha;

2331

LLVMValueRef i32_zero;

2332

LLVMValueRef check_mask;

2333

LLVMValueRef undef_src_val;

2334

2335

struct lp_build_mask_context mask_ctx;

2336

struct lp_type mask_type;

2337

struct lp_type blend_type;

2338

struct lp_type row_type;

2339

struct lp_type dst_type;

2340

struct lp_type ls_type;

2341

2342

unsigned char swizzle[TGSI_NUM_CHANNELS];

2343

unsigned vector_width;

2344

unsigned src_channels = TGSI_NUM_CHANNELS;

2345

unsigned dst_channels;

2346

unsigned dst_count;

2347

unsigned src_count;

2348

unsigned i, j;

2349

2350

const struct util_format_description* out_format_desc = util_format_description(out_format);

2351

2352

unsigned dst_alignment;

2353

2354

bool pad_inline = is_arithmetic_format(out_format_desc);

2355

bool has_alpha = false;

2356

const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&

2357

util_blend_state_is_dual(&variant->key.blend, 0);

2358

2359

const boolean is_1d = variant->key.resource_1d;

2360

boolean twiddle_after_convert = FALSE;

2361

unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;

2362

LLVMValueRef fpstate = 0;

2363

2364

/* Get type from output format */

2365

lp_blend_type_from_format_desc(out_format_desc, &row_type);

2366

lp_mem_type_from_format_desc(out_format_desc, &dst_type);

2367

2368

2369

* Technically this code should go into lp_build_smallfloat_to_float

2370

* and lp_build_float_to_smallfloat but due to the

2371

* http://llvm.org/bugs/show_bug.cgi?id=6393

2372

* llvm reorders the mxcsr intrinsics in a way that breaks the code.

2373

* So the ordering is important here and there shouldn't be any

2374

* llvm ir instrunctions in this function before

2375

* this, otherwise half-float format conversions won't work

2376

* (again due to llvm bug #6393).

2377

2378

if (have_smallfloat_format(dst_type, out_format)) {

2379

/* We need to make sure that denorms are ok for half float

2380

conversions */

2381

fpstate = lp_build_fpstate_get(gallivm);

2382

lp_build_fpstate_set_denorms_zero(gallivm, FALSE);

2383

}

2384

2385

mask_type = lp_int32_vec4_type();

2386

mask_type.length = fs_type.length;

2387

2388

for (i = num_fs; i < num_fullblock_fs; i++) {

2389

fs_mask[i] = lp_build_zero(gallivm, mask_type);

2390

}

2391

2392

/* Do not bother executing code when mask is empty.. */

2393

if (do_branch) {

2394

check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));

2395

2396

for (i = 0; i < num_fullblock_fs; ++i) {

2397

check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");

2398

}

2399

2400

lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);

2401

lp_build_mask_check(&mask_ctx);

2402

}

2403

2404

partial_mask |= !variant->opaque;

2405

i32_zero = lp_build_const_int32(gallivm, 0);

2406

2407

undef_src_val = lp_build_undef(gallivm, fs_type);

2408

2409

row_type.length = fs_type.length;

2410

vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;

2411

2412

/* Compute correct swizzle and count channels */

2413

memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);

2414

dst_channels = 0;

2415

2416

for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {

2417

/* Ensure channel is used */

2418

if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {

2419

continue;

2420

}

2421

2422

/* Ensure not already written to (happens in case with GL_ALPHA) */

2423

if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {

2424

continue;

2425

}

2426

2427

/* Ensure we haven't already found all channels */

2428

if (dst_channels >= out_format_desc->nr_channels) {

2429

continue;

2430

}

2431

2432

swizzle[out_format_desc->swizzle[i]] = i;

2433

++dst_channels;

2434

2435

if (i == alpha_channel) {

2436

has_alpha = true;

2437

}

2438

}

2439

2440

if (format_expands_to_float_soa(out_format_desc)) {

2441

2442

* the code above can't work for layout_other

2443

* for srgb it would sort of work but we short-circuit swizzles, etc.

2444

* as that is done as part of unpack / pack.

2445

2446

dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */

2447

has_alpha = true;

2448

swizzle[0] = 0;

2449

swizzle[1] = 1;

2450

swizzle[2] = 2;

2451

swizzle[3] = 3;

2452

pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */

2453

}

2454

2455

/* If 3 channels then pad to include alpha for 4 element transpose */

2456

if (dst_channels == 3) {

2457

assert (!has_alpha);

2458

for (i = 0; i < TGSI_NUM_CHANNELS; i++) {

2459

if (swizzle[i] > TGSI_NUM_CHANNELS)

2460

swizzle[i] = 3;

2461

}

2462

if (out_format_desc->nr_channels == 4) {

2463

dst_channels = 4;

2464

2465

* We use alpha from the color conversion, not separate one.

2466

* We had to include it for transpose, hence it will get converted

2467

* too (albeit when doing transpose after conversion, that would

2468

* no longer be the case necessarily).

2469

* (It works only with 4 channel dsts, e.g. rgbx formats, because

2470

* otherwise we really have padding, not alpha, included.)

2471

2472

has_alpha = true;

2473

}

2474

}

2475

2476

2477

* Load shader output

2478

2479

for (i = 0; i < num_fullblock_fs; ++i) {

2480

/* Always load alpha for use in blending */

2481

LLVMValueRef alpha;

2482

if (i < num_fs) {

2483

alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");

2484

}

2485

else {

2486

alpha = undef_src_val;

2487

}

2488

2489

/* Load each channel */

2490

for (j = 0; j < dst_channels; ++j) {

2491

assert(swizzle[j] < 4);

2492

if (i < num_fs) {

2493

fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");

2494

}

2495

else {

2496

fs_src[i][j] = undef_src_val;

2497

}

2498

}

2499

2500

/* If 3 channels then pad to include alpha for 4 element transpose */

2501

2502

* XXX If we include that here maybe could actually use it instead of

2503

* separate alpha for blending?

2504

* (Difficult though we actually convert pad channels, not alpha.)

2505

2506

if (dst_channels == 3 && !has_alpha) {

2507

fs_src[i][3] = alpha;

2508

}

2509

2510

/* We split the row_mask and row_alpha as we want 128bit interleave */

2511

if (fs_type.length == 8) {

2512

src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i],

2513

0, src_channels);

2514

src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i],

2515

src_channels, src_channels);

2516

2517

src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);

2518

src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,

2519

src_channels, src_channels);

2520

} else {

2521

src_mask[i] = fs_mask[i];

2522

src_alpha[i] = alpha;

2523

}

2524

}

2525

if (dual_source_blend) {

2526

/* same as above except different src/dst, skip masks and comments... */

2527

for (i = 0; i < num_fullblock_fs; ++i) {

2528

LLVMValueRef alpha;

2529

if (i < num_fs) {

2530

alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");

2531

}

2532

else {

2533

alpha = undef_src_val;

2534

}

2535

2536

for (j = 0; j < dst_channels; ++j) {

2537

assert(swizzle[j] < 4);

2538

if (i < num_fs) {

2539

fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");

2540

}

2541

else {

2542

fs_src1[i][j] = undef_src_val;

2543

}

2544

}

2545

if (dst_channels == 3 && !has_alpha) {

2546

fs_src1[i][3] = alpha;

2547

}

2548

if (fs_type.length == 8) {

2549

src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);

2550

src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha,

2551

src_channels, src_channels);

2552

} else {

2553

src1_alpha[i] = alpha;

2554

}

2555

}

2556

}

2557

2558

if (util_format_is_pure_integer(out_format)) {

2559

2560

* In this case fs_type was really ints or uints disguised as floats,

2561

* fix that up now.

2562

2563

fs_type.floating = 0;

2564

fs_type.sign = dst_type.sign;

2565

for (i = 0; i < num_fullblock_fs; ++i) {

2566

for (j = 0; j < dst_channels; ++j) {

2567

fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],

2568

lp_build_vec_type(gallivm, fs_type), "");

2569

}

2570

if (dst_channels == 3 && !has_alpha) {

2571

fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],

2572

lp_build_vec_type(gallivm, fs_type), "");

2573

}

2574

}

2575

}

2576

2577

2578

* We actually should generally do conversion first (for non-1d cases)

2579

* when the blend format is 8 or 16 bits. The reason is obvious,

2580

* there's 2 or 4 times less vectors to deal with for the interleave...

2581

* Albeit for the AVX (not AVX2) case there's no benefit with 16 bit

2582

* vectors (as it can do 32bit unpack with 256bit vectors, but 8/16bit

2583

* unpack only with 128bit vectors).

2584

* Note: for 16bit sizes really need matching pack conversion code

2585

2586

if (!is_1d && dst_channels != 3 && dst_type.width == 8) {

2587

twiddle_after_convert = TRUE;

2588

}

2589

2590

2591

* Pixel twiddle from fragment shader order to memory order

2592

2593

if (!twiddle_after_convert) {

2594

src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,

2595

dst_channels, fs_src, src, pad_inline);

2596

if (dual_source_blend) {

2597

generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,

2598

fs_src1, src1, pad_inline);

2599

}

2600

} else {

2601

src_count = num_fullblock_fs * dst_channels;

2602

2603

* We reorder things a bit here, so the cases for 4-wide and 8-wide

2604

* (AVX) turn out the same later when untwiddling/transpose (albeit

2605

* for true AVX2 path untwiddle needs to be different).

2606

* For now just order by colors first (so we can use unpack later).

2607

2608

for (j = 0; j < num_fullblock_fs; j++) {

2609

for (i = 0; i < dst_channels; i++) {

2610

src[i*num_fullblock_fs + j] = fs_src[j][i];

2611

if (dual_source_blend) {

2612

src1[i*num_fullblock_fs + j] = fs_src1[j][i];

2613

}

2614

}

2615

}

2616

}

2617

2618

src_channels = dst_channels < 3 ? dst_channels : 4;

2619

if (src_count != num_fullblock_fs * src_channels) {

2620

unsigned ds = src_count / (num_fullblock_fs * src_channels);

2621

row_type.length /= ds;

2622

fs_type.length = row_type.length;

2623

}

2624

2625

blend_type = row_type;

2626

mask_type.length = 4;

2627

2628

/* Convert src to row_type */

2629

if (dual_source_blend) {

2630

struct lp_type old_row_type = row_type;

2631

lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);

2632

src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);

2633

}

2634

else {

2635

src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);

2636

}

2637

2638

/* If the rows are not an SSE vector, combine them to become SSE size! */

2639

if ((row_type.width * row_type.length) % 128) {

2640

unsigned bits = row_type.width * row_type.length;

2641

unsigned combined;

2642

2643

assert(src_count >= (vector_width / bits));

2644

2645

dst_count = src_count / (vector_width / bits);

2646

2647

combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);

2648

if (dual_source_blend) {

2649

lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);

2650

}

2651

2652

row_type.length *= combined;

2653

src_count /= combined;

2654

2655

bits = row_type.width * row_type.length;

2656

assert(bits == 128 || bits == 256);

2657

}

2658

2659

if (twiddle_after_convert) {

2660

fs_twiddle_transpose(gallivm, row_type, src, src_count, src);

2661

if (dual_source_blend) {

2662

fs_twiddle_transpose(gallivm, row_type, src1, src_count, src1);

2663

}

2664

}

2665

2666

2667

* Blend Colour conversion

2668

2669

blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);

2670

blend_color = LLVMBuildPointerCast(builder, blend_color,

2671

LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");

2672

blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color,

2673

&i32_zero, 1, ""), "");

2674

2675

/* Convert */

2676

lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);

2677

2678

if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {

2679

2680

* since blending is done with floats, there was no conversion.

2681

* However, the rules according to fixed point renderbuffers still

2682

* apply, that is we must clamp inputs to 0.0/1.0.

2683

* (This would apply to separate alpha conversion too but we currently

2684

* force has_alpha to be true.)

2685

* TODO: should skip this with "fake" blend, since post-blend conversion

2686

* will clamp anyway.

2687

* TODO: could also skip this if fragment color clamping is enabled. We

2688

* don't support it natively so it gets baked into the shader however, so

2689

* can't really tell here.

2690

2691

struct lp_build_context f32_bld;

2692

assert(row_type.floating);

2693

lp_build_context_init(&f32_bld, gallivm, row_type);

2694

for (i = 0; i < src_count; i++) {

2695

src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);

2696

}

2697

if (dual_source_blend) {

2698

for (i = 0; i < src_count; i++) {

2699

src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);

2700

}

2701

}

2702

/* probably can't be different than row_type but better safe than sorry... */

2703

lp_build_context_init(&f32_bld, gallivm, blend_type);

2704

blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);

2705

}

2706

2707

/* Extract alpha */

2708

blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));

2709

2710

/* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */

2711

pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;

2712

if (pad_inline) {

2713

/* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */

2714

blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);

2715

} else {

2716

/* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */

2717

blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);

2718

}

2719

2720

2721

* Mask conversion

2722

2723

lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);

2724

2725

if (src_count < block_height) {

2726

lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);

2727

} else if (src_count > block_height) {

2728

for (i = src_count; i > 0; --i) {

2729

unsigned pixels = block_size / src_count;

2730

unsigned idx = i - 1;

2731

2732

src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],

2733

(idx * pixels) % 4, pixels);

2734

}

2735

}

2736

2737

assert(mask_type.width == 32);

2738

2739

for (i = 0; i < src_count; ++i) {

2740

unsigned pixels = block_size / src_count;

2741

unsigned pixel_width = row_type.width * dst_channels;

2742

2743

if (pixel_width == 24) {

2744

mask_type.width = 8;

2745

mask_type.length = vector_width / mask_type.width;

2746

} else {

2747

mask_type.length = pixels;

2748

mask_type.width = row_type.width * dst_channels;

2749

2750

2751

* If mask_type width is smaller than 32bit, this doesn't quite

2752

* generate the most efficient code (could use some pack).

2753

2754

src_mask[i] = LLVMBuildIntCast(builder, src_mask[i],

2755

lp_build_int_vec_type(gallivm, mask_type), "");

2756

2757

mask_type.length *= dst_channels;

2758

mask_type.width /= dst_channels;

2759

}

2760

2761

src_mask[i] = LLVMBuildBitCast(builder, src_mask[i],

2762

lp_build_int_vec_type(gallivm, mask_type), "");

2763

src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);

2764

}

2765

2766

2767

* Alpha conversion

2768

2769

if (!has_alpha) {

2770

struct lp_type alpha_type = fs_type;

2771

alpha_type.length = 4;

2772

convert_alpha(gallivm, row_type, alpha_type,

2773

block_size, block_height,

2774

src_count, dst_channels,

2775

pad_inline, src_alpha);

2776

if (dual_source_blend) {

2777

convert_alpha(gallivm, row_type, alpha_type,

2778

block_size, block_height,

2779

src_count, dst_channels,

2780

pad_inline, src1_alpha);

2781

}

2782

}

2783

2784

2785

2786

* Load dst from memory

2787

2788

if (src_count < block_height) {

2789

dst_count = block_height;

2790

} else {

2791

dst_count = src_count;

2792

}

2793

2794

dst_type.length *= block_size / dst_count;

2795

2796

if (format_expands_to_float_soa(out_format_desc)) {

2797

2798

* we need multiple values at once for the conversion, so can as well

2799

* load them vectorized here too instead of concatenating later.

2800

* (Still need concatenation later for 8-wide vectors).

2801

2802

dst_count = block_height;

2803

dst_type.length = block_width;

2804

}

2805

2806

2807

* Compute the alignment of the destination pointer in bytes

2808

* We fetch 1-4 pixels, if the format has pot alignment then those fetches

2809

* are always aligned by MIN2(16, fetch_width) except for buffers (not

2810

* 1d tex but can't distinguish here) so need to stick with per-pixel

2811

* alignment in this case.

2812

2813

if (is_1d) {

2814

dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);

2815

}

2816

else {

2817

dst_alignment = dst_type.length * dst_type.width / 8;

2818

}

2819

/* Force power-of-two alignment by extracting only the least-significant-bit */

2820

dst_alignment = 1 << (ffs(dst_alignment) - 1);

2821

2822

* Resource base and stride pointers are aligned to 16 bytes, so that's

2823

* the maximum alignment we can guarantee

2824

2825

dst_alignment = MIN2(16, dst_alignment);

2826

2827

ls_type = dst_type;

2828

2829

if (dst_count > src_count) {

2830

if ((dst_type.width == 8 || dst_type.width == 16) &&

2831

util_is_power_of_two_or_zero(dst_type.length) &&

2832

dst_type.length * dst_type.width < 128) {

2833

2834

* Never try to load values as 4xi8 which we will then

2835

* concatenate to larger vectors. This gives llvm a real

2836

* headache (the problem is the type legalizer (?) will

2837

* try to load that as 4xi8 zext to 4xi32 to fill the vector,

2838

* then the shuffles to concatenate are more or less impossible

2839

* - llvm is easily capable of generating a sequence of 32

2840

* pextrb/pinsrb instructions for that. Albeit it appears to

2841

* be fixed in llvm 4.0. So, load and concatenate with 32bit

2842

* width to avoid the trouble (16bit seems not as bad, llvm

2843

* probably recognizes the load+shuffle as only one shuffle

2844

* is necessary, but we can do just the same anyway).

2845

2846

ls_type.length = dst_type.length * dst_type.width / 32;

2847

ls_type.width = 32;

2848

}

2849

}

2850

2851

if (is_1d) {

2852

load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,

2853

dst, ls_type, dst_count / 4, dst_alignment);

2854

for (i = dst_count / 4; i < dst_count; i++) {

2855

dst[i] = lp_build_undef(gallivm, ls_type);

2856

}

2857

2858

}

2859

else {

2860

load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,

2861

dst, ls_type, dst_count, dst_alignment);

2862

}

2863

2864

2865

2866

* Convert from dst/output format to src/blending format.

2867

2868

* This is necessary as we can only read 1 row from memory at a time,

2869

* so the minimum dst_count will ever be at this point is 4.

2870

2871

* With, for example, R8 format you can have all 16 pixels in a 128 bit vector,

2872

* this will take the 4 dsts and combine them into 1 src so we can perform blending

2873

* on all 16 pixels in that single vector at once.

2874

2875

if (dst_count > src_count) {

2876

if (ls_type.length != dst_type.length && ls_type.length == 1) {

2877

LLVMTypeRef elem_type = lp_build_elem_type(gallivm, ls_type);

2878

LLVMTypeRef ls_vec_type = LLVMVectorType(elem_type, 1);

2879

for (i = 0; i < dst_count; i++) {

2880

dst[i] = LLVMBuildBitCast(builder, dst[i], ls_vec_type, "");

2881

}

2882

}

2883

2884

lp_build_concat_n(gallivm, ls_type, dst, 4, dst, src_count);

2885

2886

if (ls_type.length != dst_type.length) {

2887

struct lp_type tmp_type = dst_type;

2888

tmp_type.length = dst_type.length * 4 / src_count;

2889

for (i = 0; i < src_count; i++) {

2890

dst[i] = LLVMBuildBitCast(builder, dst[i],

2891

lp_build_vec_type(gallivm, tmp_type), "");

2892

}

2893

}

2894

}

2895

2896

2897

* Blending

2898

2899

/* XXX this is broken for RGB8 formats -

2900

* they get expanded from 12 to 16 elements (to include alpha)

2901

* by convert_to_blend_type then reduced to 15 instead of 12

2902

* by convert_from_blend_type (a simple fix though breaks A8...).

2903

* R16G16B16 also crashes differently however something going wrong

2904

* inside llvm handling npot vector sizes seemingly.

2905

* It seems some cleanup could be done here (like skipping conversion/blend

2906

* when not needed).

2907

2908

convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type,

2909

row_type, dst, src_count);

2910

2911

2912

* FIXME: Really should get logic ops / masks out of generic blend / row

2913

* format. Logic ops will definitely not work on the blend float format

2914

* used for SRGB here and I think OpenGL expects this to work as expected

2915

* (that is incoming values converted to srgb then logic op applied).

2916

2917

for (i = 0; i < src_count; ++i) {

2918

dst[i] = lp_build_blend_aos(gallivm,

2919

&variant->key.blend,

2920

out_format,

2921

row_type,

2922

rt,

2923

src[i],

2924

has_alpha ? NULL : src_alpha[i],

2925

src1[i],

2926

has_alpha ? NULL : src1_alpha[i],

2927

dst[i],

2928

partial_mask ? src_mask[i] : NULL,

2929

blend_color,

2930

has_alpha ? NULL : blend_alpha,

2931

swizzle,

2932

pad_inline ? 4 : dst_channels);

2933

}

2934

2935

convert_from_blend_type(gallivm, block_size, out_format_desc,

2936

row_type, dst_type, dst, src_count);

2937

2938

/* Split the blend rows back to memory rows */

2939

if (dst_count > src_count) {

2940

row_type.length = dst_type.length * (dst_count / src_count);

2941

2942

if (src_count == 1) {

2943

dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);

2944

dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);

2945

2946

row_type.length /= 2;

2947

src_count *= 2;

2948

}

2949

2950

dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);

2951

dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);

2952

dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);

2953

dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);

2954

2955

row_type.length /= 2;

2956

src_count *= 2;

2957

}

2958

2959

2960

* Store blend result to memory

2961

2962

if (is_1d) {

2963

store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,

2964

dst, dst_type, dst_count / 4, dst_alignment);

2965

}

2966

else {

2967

store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,

2968

dst, dst_type, dst_count, dst_alignment);

2969

}

2970

2971

if (have_smallfloat_format(dst_type, out_format)) {

2972

lp_build_fpstate_set(gallivm, fpstate);

2973

}

2974

2975

if (do_branch) {

2976

lp_build_mask_end(&mask_ctx);

2977

}

2978

}

2979

2980

2981

/**

2982

* Generate the runtime callable function for the whole fragment pipeline.

2983

* Note that the function which we generate operates on a block of 16

2984

* pixels at at time. The block contains 2x2 quads. Each quad contains

2985

* 2x2 pixels.

2986

2987

static void

2988

generate_fragment(struct llvmpipe_context *lp,

2989

struct lp_fragment_shader *shader,

2990

struct lp_fragment_shader_variant *variant,

2991

unsigned partial_mask)

2992

{

2993

struct gallivm_state *gallivm = variant->gallivm;

2994

struct lp_fragment_shader_variant_key *key = &variant->key;

2995

struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];

2996

char func_name[64];

2997

struct lp_type fs_type;

2998

struct lp_type blend_type;

2999

LLVMTypeRef fs_elem_type;

3000

LLVMTypeRef blend_vec_type;

3001

LLVMTypeRef arg_types[15];

3002

LLVMTypeRef func_type;

3003

LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);

3004

LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);

3005

LLVMValueRef context_ptr;

3006

LLVMValueRef x;

3007

LLVMValueRef y;

3008

LLVMValueRef a0_ptr;

3009

LLVMValueRef dadx_ptr;

3010

LLVMValueRef dady_ptr;

3011

LLVMValueRef color_ptr_ptr;

3012

LLVMValueRef stride_ptr;

3013

LLVMValueRef color_sample_stride_ptr;

3014

LLVMValueRef depth_ptr;

3015

LLVMValueRef depth_stride;

3016

LLVMValueRef depth_sample_stride;

3017

LLVMValueRef mask_input;

3018

LLVMValueRef thread_data_ptr;

3019

LLVMBasicBlockRef block;

3020

LLVMBuilderRef builder;

3021

struct lp_build_sampler_soa *sampler;

3022

struct lp_build_image_soa *image;

3023

struct lp_build_interp_soa_context interp;

3024

LLVMValueRef fs_mask[(16 / 4) * LP_MAX_SAMPLES];

3025

LLVMValueRef fs_out_color[LP_MAX_SAMPLES][PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];

3026

LLVMValueRef function;

3027

LLVMValueRef facing;

3028

unsigned num_fs;

3029

unsigned i;

3030

unsigned chan;

3031

unsigned cbuf;

3032

boolean cbuf0_write_all;

3033

const boolean dual_source_blend = key->blend.rt[0].blend_enable &&

3034

util_blend_state_is_dual(&key->blend, 0);

3035

3036

assert(lp_native_vector_width / 32 >= 4);

3037

3038

/* Adjust color input interpolation according to flatshade state:

3039

3040

memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);

3041

for (i = 0; i < shader->info.base.num_inputs; i++) {

3042

if (inputs[i].interp == LP_INTERP_COLOR) {

3043

if (key->flatshade)

3044

inputs[i].interp = LP_INTERP_CONSTANT;

3045

else

3046

inputs[i].interp = LP_INTERP_PERSPECTIVE;

3047

}

3048

}

3049

3050

/* check if writes to cbuf[0] are to be copied to all cbufs */

3051

cbuf0_write_all =

3052

shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];

3053

3054

/* TODO: actually pick these based on the fs and color buffer

3055

* characteristics. */

3056

3057

memset(&fs_type, 0, sizeof fs_type);

3058

fs_type.floating = TRUE; /* floating point values */

3059

fs_type.sign = TRUE; /* values are signed */

3060

fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */

3061

fs_type.width = 32; /* 32-bit float */

3062

fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */

3063

3064

memset(&blend_type, 0, sizeof blend_type);

3065

blend_type.floating = FALSE; /* values are integers */

3066

blend_type.sign = FALSE; /* values are unsigned */

3067

blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */

3068

blend_type.width = 8; /* 8-bit ubyte values */

3069

blend_type.length = 16; /* 16 elements per vector */

3070

3071

3072

* Generate the function prototype. Any change here must be reflected in

3073

* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.

3074

3075

3076

fs_elem_type = lp_build_elem_type(gallivm, fs_type);

3077

3078

blend_vec_type = lp_build_vec_type(gallivm, blend_type);

3079

3080

snprintf(func_name, sizeof(func_name), "fs_variant_%s",

3081

partial_mask ? "partial" : "whole");

3082

3083

arg_types[0] = variant->jit_context_ptr_type; /* context */

3084

arg_types[1] = int32_type; /* x */

3085

arg_types[2] = int32_type; /* y */

3086

arg_types[3] = int32_type; /* facing */

3087

arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */

3088

arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */

3089

arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */

3090

arg_types[7] = LLVMPointerType(LLVMPointerType(int8_type, 0), 0); /* color */

3091

arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */

3092

arg_types[9] = LLVMInt64TypeInContext(gallivm->context); /* mask_input */

3093

arg_types[10] = variant->jit_thread_data_ptr_type; /* per thread data */

3094

arg_types[11] = LLVMPointerType(int32_type, 0); /* stride */

3095

arg_types[12] = int32_type; /* depth_stride */

3096

arg_types[13] = LLVMPointerType(int32_type, 0); /* color sample strides */

3097

arg_types[14] = int32_type; /* depth sample stride */

3098

3099

func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),

3100

arg_types, ARRAY_SIZE(arg_types), 0);

3101

3102

function = LLVMAddFunction(gallivm->module, func_name, func_type);

3103

LLVMSetFunctionCallConv(function, LLVMCCallConv);

3104

3105

variant->function[partial_mask] = function;

3106

3107

/* XXX: need to propagate noalias down into color param now we are

3108

* passing a pointer-to-pointer?

3109

3110

for(i = 0; i < ARRAY_SIZE(arg_types); ++i)

3111

if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)

3112

lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);

3113

3114

if (variant->gallivm->cache->data_size)

3115

return;

3116

3117

context_ptr = LLVMGetParam(function, 0);

3118

x = LLVMGetParam(function, 1);

3119

y = LLVMGetParam(function, 2);

3120

facing = LLVMGetParam(function, 3);

3121

a0_ptr = LLVMGetParam(function, 4);

3122

dadx_ptr = LLVMGetParam(function, 5);

3123

dady_ptr = LLVMGetParam(function, 6);

3124

color_ptr_ptr = LLVMGetParam(function, 7);

3125

depth_ptr = LLVMGetParam(function, 8);

3126

mask_input = LLVMGetParam(function, 9);

3127

thread_data_ptr = LLVMGetParam(function, 10);

3128

stride_ptr = LLVMGetParam(function, 11);

3129

depth_stride = LLVMGetParam(function, 12);

3130

color_sample_stride_ptr = LLVMGetParam(function, 13);

3131

depth_sample_stride = LLVMGetParam(function, 14);

3132

3133

lp_build_name(context_ptr, "context");

3134

lp_build_name(x, "x");

3135

lp_build_name(y, "y");

3136

lp_build_name(a0_ptr, "a0");

3137

lp_build_name(dadx_ptr, "dadx");

3138

lp_build_name(dady_ptr, "dady");

3139

lp_build_name(color_ptr_ptr, "color_ptr_ptr");

3140

lp_build_name(depth_ptr, "depth");

3141

lp_build_name(mask_input, "mask_input");

3142

lp_build_name(thread_data_ptr, "thread_data");

3143

lp_build_name(stride_ptr, "stride_ptr");

3144

lp_build_name(depth_stride, "depth_stride");

3145

lp_build_name(color_sample_stride_ptr, "color_sample_stride_ptr");

3146

lp_build_name(depth_sample_stride, "depth_sample_stride");

3147

3148

3149

* Function body

3150

3151

3152

block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");

3153

builder = gallivm->builder;

3154

assert(builder);

3155

LLVMPositionBuilderAtEnd(builder, block);

3156

3157

3158

* Must not count ps invocations if there's a null shader.

3159

* (It would be ok to count with null shader if there's d/s tests,

3160

* but only if there's d/s buffers too, which is different

3161

* to implicit rasterization disable which must not depend

3162

* on the d/s buffers.)

3163

* Could use popcount on mask, but pixel accuracy is not required.

3164

* Could disable if there's no stats query, but maybe not worth it.

3165

3166

if (shader->info.base.num_instructions > 1) {

3167

LLVMValueRef invocs, val;

3168

invocs = lp_jit_thread_data_invocations(gallivm, thread_data_ptr);

3169

val = LLVMBuildLoad(builder, invocs, "");

3170

val = LLVMBuildAdd(builder, val,

3171

LLVMConstInt(LLVMInt64TypeInContext(gallivm->context), 1, 0),

3172

"invoc_count");

3173

LLVMBuildStore(builder, val, invocs);

3174

}

3175

3176

/* code generated texture sampling */

3177

sampler = lp_llvm_sampler_soa_create(lp_fs_variant_key_samplers(key), key->nr_samplers);

3178

image = lp_llvm_image_soa_create(lp_fs_variant_key_images(key), key->nr_images);

3179

3180

num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */

3181

/* for 1d resources only run "upper half" of stamp */

3182

if (key->resource_1d)

3183

num_fs /= 2;

3184

3185

{

3186

LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);

3187

LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);

3188

LLVMValueRef num_loop_samp = lp_build_const_int32(gallivm, num_fs * key->coverage_samples);

3189

LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,

3190

num_loop_samp, "mask_store");

3191

3192

LLVMTypeRef flt_type = LLVMFloatTypeInContext(gallivm->context);

3193

LLVMValueRef glob_sample_pos = LLVMAddGlobal(gallivm->module, LLVMArrayType(flt_type, key->coverage_samples * 2), "");

3194

LLVMValueRef sample_pos_array;

3195

3196

if (key->multisample && key->coverage_samples == 4) {

3197

LLVMValueRef sample_pos_arr[8];

3198

for (unsigned i = 0; i < 4; i++) {

3199

sample_pos_arr[i * 2] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][0]);

3200

sample_pos_arr[i * 2 + 1] = LLVMConstReal(flt_type, lp_sample_pos_4x[i][1]);

3201

}

3202

sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 8);

3203

} else {

3204

LLVMValueRef sample_pos_arr[2];

3205

sample_pos_arr[0] = LLVMConstReal(flt_type, 0.5);

3206

sample_pos_arr[1] = LLVMConstReal(flt_type, 0.5);

3207

sample_pos_array = LLVMConstArray(LLVMFloatTypeInContext(gallivm->context), sample_pos_arr, 2);

3208

}

3209

LLVMSetInitializer(glob_sample_pos, sample_pos_array);

3210

3211

LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];

3212

boolean pixel_center_integer =

3213

shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];

3214

3215

3216

* The shader input interpolation info is not explicitely baked in the

3217

* shader key, but everything it derives from (TGSI, and flatshade) is

3218

* already included in the shader key.

3219

3220

lp_build_interp_soa_init(&interp,

3221

gallivm,

3222

shader->info.base.num_inputs,

3223

inputs,

3224

pixel_center_integer,

3225

key->coverage_samples, glob_sample_pos,

3226

num_loop,

3227

key->depth_clamp,

3228

builder, fs_type,

3229

a0_ptr, dadx_ptr, dady_ptr,

3230

x, y);

3231

3232

for (i = 0; i < num_fs; i++) {

3233

if (key->multisample) {

3234

LLVMValueRef smask_val = LLVMBuildLoad(builder, lp_jit_context_sample_mask(gallivm, context_ptr), "");

3235

3236

3237

* For multisampling, extract the per-sample mask from the incoming 64-bit mask,

3238

* store to the per sample mask storage. Or all of them together to generate

3239

* the fragment shader mask. (sample shading TODO).

3240

* Take the incoming state coverage mask into account.

3241

3242

for (unsigned s = 0; s < key->coverage_samples; s++) {

3243

LLVMValueRef sindexi = lp_build_const_int32(gallivm, i + (s * num_fs));

3244

LLVMValueRef sample_mask_ptr = LLVMBuildGEP(builder, mask_store,

3245

&sindexi, 1, "sample_mask_ptr");

3246

LLVMValueRef s_mask = generate_quad_mask(gallivm, fs_type,

3247

i*fs_type.length/4, s, mask_input);

3248

3249

LLVMValueRef smask_bit = LLVMBuildAnd(builder, smask_val, lp_build_const_int32(gallivm, (1 << s)), "");

3250

LLVMValueRef cmp = LLVMBuildICmp(builder, LLVMIntNE, smask_bit, lp_build_const_int32(gallivm, 0), "");

3251

smask_bit = LLVMBuildSExt(builder, cmp, int32_type, "");

3252

smask_bit = lp_build_broadcast(gallivm, mask_type, smask_bit);

3253

3254

s_mask = LLVMBuildAnd(builder, s_mask, smask_bit, "");

3255

LLVMBuildStore(builder, s_mask, sample_mask_ptr);

3256

}

3257

} else {

3258

LLVMValueRef mask;

3259

LLVMValueRef indexi = lp_build_const_int32(gallivm, i);

3260

LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,

3261

&indexi, 1, "mask_ptr");

3262

3263

if (partial_mask) {

3264

mask = generate_quad_mask(gallivm, fs_type,

3265

i*fs_type.length/4, 0, mask_input);

3266

}

3267

else {

3268

mask = lp_build_const_int_vec(gallivm, fs_type, ~0);

3269

}

3270

LLVMBuildStore(builder, mask, mask_ptr);

3271

}

3272

}

3273

3274

generate_fs_loop(gallivm,

3275

shader, key,

3276

builder,

3277

fs_type,

3278

context_ptr,

3279

glob_sample_pos,

3280

num_loop,

3281

&interp,

3282

sampler,

3283

image,

3284

mask_store, /* output */

3285

color_store,

3286

depth_ptr,

3287

depth_stride,

3288

depth_sample_stride,

3289

color_ptr_ptr,

3290

stride_ptr,

3291

color_sample_stride_ptr,

3292

facing,

3293

thread_data_ptr);

3294

3295

for (i = 0; i < num_fs; i++) {

3296

LLVMValueRef ptr;

3297

for (unsigned s = 0; s < key->coverage_samples; s++) {

3298

int idx = (i + (s * num_fs));

3299

LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);

3300

ptr = LLVMBuildGEP(builder, mask_store, &sindexi, 1, "");

3301

3302

fs_mask[idx] = LLVMBuildLoad(builder, ptr, "smask");

3303

}

3304

3305

for (unsigned s = 0; s < key->min_samples; s++) {

3306

/* This is fucked up need to reorganize things */

3307

int idx = s * num_fs + i;

3308

LLVMValueRef sindexi = lp_build_const_int32(gallivm, idx);

3309

for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {

3310

for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {

3311

ptr = LLVMBuildGEP(builder,

3312

color_store[cbuf * !cbuf0_write_all][chan],

3313

&sindexi, 1, "");

3314

fs_out_color[s][cbuf][chan][i] = ptr;

3315

}

3316

}

3317

if (dual_source_blend) {

3318

/* only support one dual source blend target hence always use output 1 */

3319

for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {

3320

ptr = LLVMBuildGEP(builder,

3321

color_store[1][chan],

3322

&sindexi, 1, "");

3323

fs_out_color[s][1][chan][i] = ptr;

3324

}

3325

}

3326

}

3327

}

3328

}

3329

3330

sampler->destroy(sampler);

3331

image->destroy(image);

3332

/* Loop over color outputs / color buffers to do blending.

3333

3334

for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {

3335

if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {

3336

LLVMValueRef color_ptr;

3337

LLVMValueRef stride;

3338

LLVMValueRef sample_stride = NULL;

3339

LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);

3340

3341

boolean do_branch = ((key->depth.enabled

3342

|| key->stencil[0].enabled

3343

|| key->alpha.enabled)

3344

&& !shader->info.base.uses_kill);

3345

3346

color_ptr = LLVMBuildLoad(builder,

3347

LLVMBuildGEP(builder, color_ptr_ptr,

3348

&index, 1, ""),

3349

"");

3350

3351

stride = LLVMBuildLoad(builder,

3352

LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),

3353

"");

3354

3355

if (key->cbuf_nr_samples[cbuf] > 1)

3356

sample_stride = LLVMBuildLoad(builder,

3357

LLVMBuildGEP(builder, color_sample_stride_ptr,

3358

&index, 1, ""), "");

3359

3360

for (unsigned s = 0; s < key->cbuf_nr_samples[cbuf]; s++) {

3361

unsigned mask_idx = num_fs * (key->multisample ? s : 0);

3362

unsigned out_idx = key->min_samples == 1 ? 0 : s;

3363

LLVMValueRef out_ptr = color_ptr;;

3364

3365

if (sample_stride) {

3366

LLVMValueRef sample_offset = LLVMBuildMul(builder, sample_stride, lp_build_const_int32(gallivm, s), "");

3367

out_ptr = LLVMBuildGEP(builder, out_ptr, &sample_offset, 1, "");

3368

}

3369

out_ptr = LLVMBuildBitCast(builder, out_ptr, LLVMPointerType(blend_vec_type, 0), "");

3370

3371

lp_build_name(out_ptr, "color_ptr%d", cbuf);

3372

3373

generate_unswizzled_blend(gallivm, cbuf, variant,

3374

key->cbuf_format[cbuf],

3375

num_fs, fs_type, &fs_mask[mask_idx], fs_out_color[out_idx],

3376

context_ptr, out_ptr, stride,

3377

partial_mask, do_branch);

3378

}

3379

}

3380

}

3381

3382

LLVMBuildRetVoid(builder);

3383

3384

gallivm_verify_function(gallivm, function);

3385

}

3386

3387

3388

static void

3389

dump_fs_variant_key(struct lp_fragment_shader_variant_key *key)

3390

{

3391

unsigned i;

3392

3393

debug_printf("fs variant %p:\n", (void *) key);

3394

3395

if (key->flatshade) {

3396

debug_printf("flatshade = 1\n");

3397

}

3398

if (key->depth_clamp)

3399

debug_printf("depth_clamp = 1\n");

3400

3401

if (key->multisample) {

3402

debug_printf("multisample = 1\n");

3403

debug_printf("coverage samples = %d\n", key->coverage_samples);

3404

debug_printf("min samples = %d\n", key->min_samples);

3405

}

3406

for (i = 0; i < key->nr_cbufs; ++i) {

3407

debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));

3408

debug_printf("cbuf nr_samples[%u] = %d\n", i, key->cbuf_nr_samples[i]);

3409

}

3410

if (key->depth.enabled || key->stencil[0].enabled) {

3411

debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));

3412

debug_printf("depth nr_samples = %d\n", key->zsbuf_nr_samples);

3413

}

3414

if (key->depth.enabled) {

3415

debug_printf("depth.func = %s\n", util_str_func(key->depth.func, TRUE));

3416

debug_printf("depth.writemask = %u\n", key->depth.writemask);

3417

}

3418

3419

for (i = 0; i < 2; ++i) {

3420

if (key->stencil[i].enabled) {

3421

debug_printf("stencil[%u].func = %s\n", i, util_str_func(key->stencil[i].func, TRUE));

3422

debug_printf("stencil[%u].fail_op = %s\n", i, util_str_stencil_op(key->stencil[i].fail_op, TRUE));

3423

debug_printf("stencil[%u].zpass_op = %s\n", i, util_str_stencil_op(key->stencil[i].zpass_op, TRUE));

3424

debug_printf("stencil[%u].zfail_op = %s\n", i, util_str_stencil_op(key->stencil[i].zfail_op, TRUE));

3425

debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);

3426

debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);

3427

}

3428

}

3429

3430

if (key->alpha.enabled) {

3431

debug_printf("alpha.func = %s\n", util_str_func(key->alpha.func, TRUE));

3432

}

3433

3434

if (key->occlusion_count) {

3435

debug_printf("occlusion_count = 1\n");

3436

}

3437

3438

if (key->blend.logicop_enable) {

3439

debug_printf("blend.logicop_func = %s\n", util_str_logicop(key->blend.logicop_func, TRUE));

3440

}

3441

else if (key->blend.rt[0].blend_enable) {

3442

debug_printf("blend.rgb_func = %s\n", util_str_blend_func (key->blend.rt[0].rgb_func, TRUE));

3443

debug_printf("blend.rgb_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));

3444

debug_printf("blend.rgb_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));

3445

debug_printf("blend.alpha_func = %s\n", util_str_blend_func (key->blend.rt[0].alpha_func, TRUE));

3446

debug_printf("blend.alpha_src_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));

3447

debug_printf("blend.alpha_dst_factor = %s\n", util_str_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));

3448

}

3449

debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);

3450

if (key->blend.alpha_to_coverage) {

3451

debug_printf("blend.alpha_to_coverage is enabled\n");

3452

}

3453

for (i = 0; i < key->nr_samplers; ++i) {

3454

const struct lp_sampler_static_state *samplers = lp_fs_variant_key_samplers(key);

3455

const struct lp_static_sampler_state *sampler = &samplers[i].sampler_state;

3456

debug_printf("sampler[%u] = \n", i);

3457

debug_printf(" .wrap = %s %s %s\n",

3458

util_str_tex_wrap(sampler->wrap_s, TRUE),

3459

util_str_tex_wrap(sampler->wrap_t, TRUE),

3460

util_str_tex_wrap(sampler->wrap_r, TRUE));

3461

debug_printf(" .min_img_filter = %s\n",

3462

util_str_tex_filter(sampler->min_img_filter, TRUE));

3463

debug_printf(" .min_mip_filter = %s\n",

3464

util_str_tex_mipfilter(sampler->min_mip_filter, TRUE));

3465

debug_printf(" .mag_img_filter = %s\n",

3466

util_str_tex_filter(sampler->mag_img_filter, TRUE));

3467

if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)

3468

debug_printf(" .compare_func = %s\n", util_str_func(sampler->compare_func, TRUE));

3469

debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);

3470

debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);

3471

debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);

3472

debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);

3473

debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);

3474

debug_printf(" .reduction_mode = %u\n", sampler->reduction_mode);

3475

debug_printf(" .aniso = %u\n", sampler->aniso);

3476

}

3477

for (i = 0; i < key->nr_sampler_views; ++i) {

3478

const struct lp_sampler_static_state *samplers = lp_fs_variant_key_samplers(key);

3479

const struct lp_static_texture_state *texture = &samplers[i].texture_state;

3480

debug_printf("texture[%u] = \n", i);

3481

debug_printf(" .format = %s\n",

3482

util_format_name(texture->format));

3483

debug_printf(" .target = %s\n",

3484

util_str_tex_target(texture->target, TRUE));

3485

debug_printf(" .level_zero_only = %u\n",

3486

texture->level_zero_only);

3487

debug_printf(" .pot = %u %u %u\n",

3488

texture->pot_width,

3489

texture->pot_height,

3490

texture->pot_depth);

3491

}

3492

struct lp_image_static_state *images = lp_fs_variant_key_images(key);

3493

for (i = 0; i < key->nr_images; ++i) {

3494

const struct lp_static_texture_state *image = &images[i].image_state;

3495

debug_printf("image[%u] = \n", i);

3496

debug_printf(" .format = %s\n",

3497

util_format_name(image->format));

3498

debug_printf(" .target = %s\n",

3499

util_str_tex_target(image->target, TRUE));

3500

debug_printf(" .level_zero_only = %u\n",

3501

image->level_zero_only);

3502

debug_printf(" .pot = %u %u %u\n",

3503

image->pot_width,

3504

image->pot_height,

3505

image->pot_depth);

3506

}

3507

}

3508

3509

const char *

3510

lp_debug_fs_kind(enum lp_fs_kind kind)

3511

{

3512

switch(kind) {

3513

case LP_FS_KIND_GENERAL:

3514

return "GENERAL";

3515

case LP_FS_KIND_BLIT_RGBA:

3516

return "BLIT_RGBA";

3517

case LP_FS_KIND_BLIT_RGB1:

3518

return "BLIT_RGB1";

3519

case LP_FS_KIND_AERO_MINIFICATION:

3520

return "AERO_MINIFICATION";

3521

case LP_FS_KIND_LLVM_LINEAR:

3522

return "LLVM_LINEAR";

3523

default:

3524

return "unknown";

3525

}

3526

}

3527

3528

void

3529

lp_debug_fs_variant(struct lp_fragment_shader_variant *variant)

3530

{

3531

debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",

3532

variant->shader->no, variant->no);

3533

if (variant->shader->base.type == PIPE_SHADER_IR_TGSI)

3534

tgsi_dump(variant->shader->base.tokens, 0);

3535

else

3536

nir_print_shader(variant->shader->base.ir.nir, stderr);

3537

dump_fs_variant_key(&variant->key);

3538

debug_printf("variant->opaque = %u\n", variant->opaque);

3539

debug_printf("variant->potentially_opaque = %u\n", variant->potentially_opaque);

3540

debug_printf("variant->blit = %u\n", variant->blit);

3541

debug_printf("shader->kind = %s\n", lp_debug_fs_kind(variant->shader->kind));

3542

debug_printf("\n");

3543

}

3544

3545

static void

3546

lp_fs_get_ir_cache_key(struct lp_fragment_shader_variant *variant,

3547

unsigned char ir_sha1_cache_key[20])

3548

{

3549

struct blob blob = { 0 };

3550

unsigned ir_size;

3551

void *ir_binary;

3552

3553

blob_init(&blob);

3554

nir_serialize(&blob, variant->shader->base.ir.nir, true);

3555

ir_binary = blob.data;

3556

ir_size = blob.size;

3557

3558

struct mesa_sha1 ctx;

3559

_mesa_sha1_init(&ctx);

3560

_mesa_sha1_update(&ctx, &variant->key, variant->shader->variant_key_size);

3561

_mesa_sha1_update(&ctx, ir_binary, ir_size);

3562

_mesa_sha1_final(&ctx, ir_sha1_cache_key);

3563

3564

blob_finish(&blob);

3565

}

3566

3567

/**

3568

* Generate a new fragment shader variant from the shader code and

3569

* other state indicated by the key.

3570

3571

static struct lp_fragment_shader_variant *

3572

generate_variant(struct llvmpipe_context *lp,

3573

struct lp_fragment_shader *shader,

3574

const struct lp_fragment_shader_variant_key *key)

3575

{

3576

struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);

3577

struct lp_fragment_shader_variant *variant;

3578

const struct util_format_description *cbuf0_format_desc = NULL;

3579

boolean fullcolormask;

3580

boolean no_kill;

3581

boolean linear;

3582

char module_name[64];

3583

unsigned char ir_sha1_cache_key[20];

3584

struct lp_cached_code cached = { 0 };

3585

bool needs_caching = false;

3586

variant = MALLOC(sizeof *variant + shader->variant_key_size - sizeof variant->key);

3587

if (!variant)

3588

return NULL;

3589

3590

memset(variant, 0, sizeof(*variant));

3591

snprintf(module_name, sizeof(module_name), "fs%u_variant%u",

3592

shader->no, shader->variants_created);

3593

3594

pipe_reference_init(&variant->reference, 1);

3595

lp_fs_reference(lp, &variant->shader, shader);

3596

3597

memcpy(&variant->key, key, shader->variant_key_size);

3598

3599

if (shader->base.ir.nir) {

3600

lp_fs_get_ir_cache_key(variant, ir_sha1_cache_key);

3601

3602

lp_disk_cache_find_shader(screen, &cached, ir_sha1_cache_key);

3603

if (!cached.data_size)

3604

needs_caching = true;

3605

}

3606

variant->gallivm = gallivm_create(module_name, lp->context, &cached);

3607

if (!variant->gallivm) {

3608

FREE(variant);

3609

return NULL;

3610

}

3611

3612

variant->list_item_global.base = variant;

3613

variant->list_item_local.base = variant;

3614

variant->no = shader->variants_created++;

3615

3616

3617

3618

3619

* Determine whether we are touching all channels in the color buffer.

3620

3621

fullcolormask = FALSE;

3622

if (key->nr_cbufs == 1) {

3623

cbuf0_format_desc = util_format_description(key->cbuf_format[0]);

3624

fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);

3625

}

3626

3627

/* The scissor is ignored here as only tiles inside the scissoring

3628

* rectangle will refer to this */

3629

no_kill =

3630

fullcolormask &&

3631

!key->stencil[0].enabled &&

3632

!key->alpha.enabled &&

3633

!key->multisample &&

3634

!key->blend.alpha_to_coverage &&

3635

!key->depth.enabled &&

3636

!shader->info.base.uses_kill &&

3637

!shader->info.base.writes_samplemask &&

3638

!shader->info.base.uses_fbfetch;

3639

3640

variant->opaque =

3641

no_kill &&

3642

!key->blend.logicop_enable &&

3643

!key->blend.rt[0].blend_enable

3644

? TRUE : FALSE;

3645

3646

variant->potentially_opaque =

3647

no_kill &&

3648

!key->blend.logicop_enable &&

3649

key->blend.rt[0].blend_enable &&

3650

key->blend.rt[0].rgb_func == PIPE_BLEND_ADD &&

3651

key->blend.rt[0].rgb_dst_factor == PIPE_BLENDFACTOR_INV_SRC_ALPHA &&

3652

key->blend.rt[0].alpha_func == key->blend.rt[0].rgb_func &&

3653

key->blend.rt[0].alpha_dst_factor == key->blend.rt[0].rgb_dst_factor &&

3654

shader->base.type == PIPE_SHADER_IR_TGSI &&

3655

3656

* FIXME: for NIR, all of the fields of info.xxx (except info.base)

3657

* are zeros, hence shader analysis (here and elsewhere) using these

3658

* bits cannot work and will silently fail (cbuf is the only pointer

3659

* field, hence causing a crash).

3660

3661

shader->info.cbuf[0][3].file != TGSI_FILE_NULL

3662

? TRUE : FALSE;

3663

3664

/* We only care about opaque blits for now */

3665

if (variant->opaque &&

3666

(shader->kind == LP_FS_KIND_BLIT_RGBA ||

3667

shader->kind == LP_FS_KIND_BLIT_RGB1)) {

3668

unsigned target, min_img_filter, mag_img_filter, min_mip_filter;

3669

enum pipe_format texture_format;

3670

struct lp_sampler_static_state *samp0 = lp_fs_variant_key_sampler_idx(key, 0);

3671

assert(samp0);

3672

texture_format = samp0->texture_state.format;

3673

target = samp0->texture_state.target;

3674

min_img_filter = samp0->sampler_state.min_img_filter;

3675

mag_img_filter = samp0->sampler_state.mag_img_filter;

3676

if (samp0->texture_state.level_zero_only) {

3677

min_mip_filter = PIPE_TEX_MIPFILTER_NONE;

3678

} else {

3679

min_mip_filter = samp0->sampler_state.min_mip_filter;

3680

}

3681

3682

if (target == PIPE_TEXTURE_2D &&

3683

min_img_filter == PIPE_TEX_FILTER_NEAREST &&

3684

mag_img_filter == PIPE_TEX_FILTER_NEAREST &&

3685

min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&

3686

((texture_format &&

3687

util_is_format_compatible(util_format_description(texture_format),

3688

cbuf0_format_desc)) ||

3689

(shader->kind == LP_FS_KIND_BLIT_RGB1 &&

3690

(texture_format == PIPE_FORMAT_B8G8R8A8_UNORM ||

3691

texture_format == PIPE_FORMAT_B8G8R8X8_UNORM) &&

3692

(key->cbuf_format[0] == PIPE_FORMAT_B8G8R8A8_UNORM ||

3693

key->cbuf_format[0] == PIPE_FORMAT_B8G8R8X8_UNORM))))

3694

variant->blit = 1;

3695

}

3696

3697

3698

/* Whether this is a candidate for the linear path */

3699

linear =

3700

!key->stencil[0].enabled &&

3701

!key->depth.enabled &&

3702

!shader->info.base.uses_kill &&

3703

!key->blend.logicop_enable &&

3704

(key->cbuf_format[0] == PIPE_FORMAT_B8G8R8A8_UNORM ||

3705

key->cbuf_format[0] == PIPE_FORMAT_B8G8R8X8_UNORM);

3706

3707

memcpy(&variant->key, key, sizeof *key);

3708

3709

if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {

3710

lp_debug_fs_variant(variant);

3711

}

3712

3713

llvmpipe_fs_variant_fastpath(variant);

3714

3715

lp_jit_init_types(variant);

3716

3717

if (variant->jit_function[RAST_EDGE_TEST] == NULL)

3718

generate_fragment(lp, shader, variant, RAST_EDGE_TEST);

3719

3720

if (variant->jit_function[RAST_WHOLE] == NULL) {

3721

if (variant->opaque) {

3722

/* Specialized shader, which doesn't need to read the color buffer. */

3723

generate_fragment(lp, shader, variant, RAST_WHOLE);

3724

}

3725

}

3726

3727

if (linear) {

3728

/* Currently keeping both the old fastpaths and new linear path

3729

* active. The older code is still somewhat faster for the cases

3730

* it covers.

3731

3732

* XXX: consider restricting this to aero-mode only.

3733

3734

if (fullcolormask &&

3735

!key->alpha.enabled &&

3736

!key->blend.alpha_to_coverage) {

3737

llvmpipe_fs_variant_linear_fastpath(variant);

3738

}

3739

3740

/* If the original fastpath doesn't cover this variant, try the new

3741

* code:

3742

3743

if (variant->jit_linear == NULL) {

3744

if (shader->kind == LP_FS_KIND_BLIT_RGBA ||

3745

shader->kind == LP_FS_KIND_BLIT_RGB1 ||

3746

shader->kind == LP_FS_KIND_LLVM_LINEAR) {

3747

llvmpipe_fs_variant_linear_llvm(lp, shader, variant);

3748

}

3749

}

3750

} else {

3751

if (LP_DEBUG & DEBUG_LINEAR) {

3752

lp_debug_fs_variant(variant);

3753

debug_printf(" ----> no linear path for this variant\n");

3754

}

3755

}

3756

3757

3758

* Compile everything

3759

3760

3761

gallivm_compile_module(variant->gallivm);

3762

3763

variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);

3764

3765

if (variant->function[RAST_EDGE_TEST]) {

3766

variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)

3767

gallivm_jit_function(variant->gallivm,

3768

variant->function[RAST_EDGE_TEST]);

3769

}

3770

3771

if (variant->function[RAST_WHOLE]) {

3772

variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)

3773

gallivm_jit_function(variant->gallivm,

3774

variant->function[RAST_WHOLE]);

3775

} else if (!variant->jit_function[RAST_WHOLE]) {

3776

variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];

3777

}

3778

3779

if (linear) {

3780

if (variant->linear_function) {

3781

variant->jit_linear_llvm = (lp_jit_linear_llvm_func)

3782

gallivm_jit_function(variant->gallivm, variant->linear_function);

3783

}

3784

3785

3786

* This must be done after LLVM compilation, as it will call the JIT'ed

3787

* code to determine active inputs.

3788

3789

lp_linear_check_variant(variant);

3790

}

3791

3792

if (needs_caching) {

3793

lp_disk_cache_insert_shader(screen, &cached, ir_sha1_cache_key);

3794

}

3795

3796

gallivm_free_ir(variant->gallivm);

3797

3798

return variant;

3799

}

3800

3801

3802

static void *

3803

llvmpipe_create_fs_state(struct pipe_context *pipe,

3804

const struct pipe_shader_state *templ)

3805

{

3806

struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);

3807

struct lp_fragment_shader *shader;

3808

int nr_samplers;

3809

int nr_sampler_views;

3810

int nr_images;

3811

int i;

3812

3813

shader = CALLOC_STRUCT(lp_fragment_shader);

3814

if (!shader)

3815

return NULL;

3816

3817

pipe_reference_init(&shader->reference, 1);

3818

shader->no = fs_no++;

3819

make_empty_list(&shader->variants);

3820

3821

shader->base.type = templ->type;

3822

if (templ->type == PIPE_SHADER_IR_TGSI) {

3823

/* get/save the summary info for this shader */

3824

lp_build_tgsi_info(templ->tokens, &shader->info);

3825

3826

/* we need to keep a local copy of the tokens */

3827

shader->base.tokens = tgsi_dup_tokens(templ->tokens);

3828

} else {

3829

shader->base.ir.nir = templ->ir.nir;

3830

nir_tgsi_scan_shader(templ->ir.nir, &shader->info.base, true);

3831

}

3832

3833

shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);

3834

if (shader->draw_data == NULL) {

3835

FREE((void *) shader->base.tokens);

3836

FREE(shader);

3837

return NULL;

3838

}

3839

3840

nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;

3841

nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;

3842

nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;

3843

shader->variant_key_size = lp_fs_variant_key_size(MAX2(nr_samplers, nr_sampler_views), nr_images);

3844

3845

for (i = 0; i < shader->info.base.num_inputs; i++) {

3846

shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];

3847

shader->inputs[i].location = shader->info.base.input_interpolate_loc[i];

3848

3849

switch (shader->info.base.input_interpolate[i]) {

3850

case TGSI_INTERPOLATE_CONSTANT:

3851

shader->inputs[i].interp = LP_INTERP_CONSTANT;

3852

break;

3853

case TGSI_INTERPOLATE_LINEAR:

3854

shader->inputs[i].interp = LP_INTERP_LINEAR;

3855

break;

3856

case TGSI_INTERPOLATE_PERSPECTIVE:

3857

shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;

3858

break;

3859

case TGSI_INTERPOLATE_COLOR:

3860

shader->inputs[i].interp = LP_INTERP_COLOR;

3861

break;

3862

default:

3863

assert(0);

3864

break;

3865

}

3866

3867

switch (shader->info.base.input_semantic_name[i]) {

3868

case TGSI_SEMANTIC_FACE:

3869

shader->inputs[i].interp = LP_INTERP_FACING;

3870

break;

3871

case TGSI_SEMANTIC_POSITION:

3872

/* Position was already emitted above

3873

3874

shader->inputs[i].interp = LP_INTERP_POSITION;

3875

shader->inputs[i].src_index = 0;

3876

continue;

3877

}

3878

3879

/* XXX this is a completely pointless index map... */

3880

shader->inputs[i].src_index = i+1;

3881

}

3882

3883

if (LP_DEBUG & DEBUG_TGSI && templ->type == PIPE_SHADER_IR_TGSI) {

3884

unsigned attrib;

3885

debug_printf("llvmpipe: Create fragment shader #%u %p:\n",

3886

shader->no, (void *) shader);

3887

tgsi_dump(templ->tokens, 0);

3888

debug_printf("usage masks:\n");

3889

for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {

3890

unsigned usage_mask = shader->info.base.input_usage_mask[attrib];

3891

debug_printf(" IN[%u].%s%s%s%s\n",

3892

attrib,

3893

usage_mask & TGSI_WRITEMASK_X ? "x" : "",

3894

usage_mask & TGSI_WRITEMASK_Y ? "y" : "",

3895

usage_mask & TGSI_WRITEMASK_Z ? "z" : "",

3896

usage_mask & TGSI_WRITEMASK_W ? "w" : "");

3897

}

3898

debug_printf("\n");

3899

}

3900

3901

/* This will put a derived copy of the tokens into shader->base.tokens */

3902

if (templ->type == PIPE_SHADER_IR_TGSI)

3903

llvmpipe_fs_analyse(shader, templ->tokens);

3904

else

3905

llvmpipe_fs_analyse_nir(shader);

3906

3907

return shader;

3908

}

3909

3910

3911

static void

3912

llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)

3913

{

3914

struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);

3915

struct lp_fragment_shader *lp_fs = (struct lp_fragment_shader *)fs;

3916

if (llvmpipe->fs == lp_fs)

3917

return;

3918

3919

draw_bind_fragment_shader(llvmpipe->draw,

3920

(lp_fs ? lp_fs->draw_data : NULL));

3921

3922

lp_fs_reference(llvmpipe, &llvmpipe->fs, lp_fs);

3923

3924

/* invalidate the setup link, NEW_FS will make it update */

3925

lp_setup_set_fs_variant(llvmpipe->setup, NULL);

3926

llvmpipe->dirty |= LP_NEW_FS;

3927

}

3928

3929

3930

/**

3931

* Remove shader variant from two lists: the shader's variant list

3932

* and the context's variant list.

3933

3934

3935

static

3936

void llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,

3937

struct lp_fragment_shader_variant *variant)

3938

{

3939

if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {

3940

debug_printf("llvmpipe: del fs #%u var %u v created %u v cached %u "

3941

"v total cached %u inst %u total inst %u\n",

3942

variant->shader->no, variant->no,

3943

variant->shader->variants_created,

3944

variant->shader->variants_cached,

3945

lp->nr_fs_variants, variant->nr_instrs, lp->nr_fs_instrs);

3946

}

3947

3948

/* remove from shader's list */

3949

remove_from_list(&variant->list_item_local);

3950

variant->shader->variants_cached--;

3951

3952

/* remove from context's list */

3953

remove_from_list(&variant->list_item_global);

3954

lp->nr_fs_variants--;

3955

lp->nr_fs_instrs -= variant->nr_instrs;

3956

}

3957

3958

void

3959

llvmpipe_destroy_shader_variant(struct llvmpipe_context *lp,

3960

struct lp_fragment_shader_variant *variant)

3961

{

3962

gallivm_destroy(variant->gallivm);

3963

3964

lp_fs_reference(lp, &variant->shader, NULL);

3965

3966

FREE(variant);

3967

}

3968

3969

void

3970

llvmpipe_destroy_fs(struct llvmpipe_context *llvmpipe,

3971

struct lp_fragment_shader *shader)

3972

{

3973

/* Delete draw module's data */

3974

draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);

3975

3976

if (shader->base.ir.nir)

3977

ralloc_free(shader->base.ir.nir);

3978

assert(shader->variants_cached == 0);

3979

FREE((void *) shader->base.tokens);

3980

FREE(shader);

3981

}

3982

3983

static void

3984

llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)

3985

{

3986

struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);

3987

struct lp_fragment_shader *shader = fs;

3988

struct lp_fs_variant_list_item *li;

3989

3990

/* Delete all the variants */

3991

li = first_elem(&shader->variants);

3992

while(!at_end(&shader->variants, li)) {

3993

struct lp_fs_variant_list_item *next = next_elem(li);

3994

struct lp_fragment_shader_variant *variant;

3995

variant = li->base;

3996

llvmpipe_remove_shader_variant(llvmpipe, li->base);

3997

lp_fs_variant_reference(llvmpipe, &variant, NULL);

3998

li = next;

3999

}

4000

4001

lp_fs_reference(llvmpipe, &shader, NULL);

4002

}

4003

4004

static void

4005

llvmpipe_set_constant_buffer(struct pipe_context *pipe,

4006

enum pipe_shader_type shader, uint index,

4007

bool take_ownership,

4008

const struct pipe_constant_buffer *cb)

4009

{

4010

struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);

4011

struct pipe_constant_buffer *constants = &llvmpipe->constants[shader][index];

4012

4013

assert(shader < PIPE_SHADER_TYPES);

4014

assert(index < ARRAY_SIZE(llvmpipe->constants[shader]));

4015

4016

/* note: reference counting */

4017

util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb,

4018

take_ownership);

4019

4020

/* user_buffer is only valid until the next set_constant_buffer (at most,

4021

* possibly until shader deletion), so we need to upload it now to make sure

4022

* it doesn't get updated/freed out from under us.

4023

4024

if (constants->user_buffer) {

4025

u_upload_data(llvmpipe->pipe.const_uploader, 0, constants->buffer_size, 16,

4026

constants->user_buffer, &constants->buffer_offset,

4027

&constants->buffer);

4028

}

4029

if (constants->buffer) {

4030

if (!(constants->buffer->bind & PIPE_BIND_CONSTANT_BUFFER)) {

4031

debug_printf("Illegal set constant without bind flag\n");

4032

constants->buffer->bind |= PIPE_BIND_CONSTANT_BUFFER;

4033

}

4034

}

4035

4036

if (shader == PIPE_SHADER_VERTEX ||

4037

shader == PIPE_SHADER_GEOMETRY ||

4038

shader == PIPE_SHADER_TESS_CTRL ||

4039

shader == PIPE_SHADER_TESS_EVAL) {

4040

/* Pass the constants to the 'draw' module */

4041

const unsigned size = cb ? cb->buffer_size : 0;

4042

4043

const ubyte *data = NULL;

4044

if (constants->buffer)

4045

data = (ubyte *) llvmpipe_resource_data(constants->buffer) + constants->buffer_offset;

4046

4047

draw_set_mapped_constant_buffer(llvmpipe->draw, shader,

4048

index, data, size);

4049

}

4050

else if (shader == PIPE_SHADER_COMPUTE)

4051

llvmpipe->cs_dirty |= LP_CSNEW_CONSTANTS;

4052

else

4053

llvmpipe->dirty |= LP_NEW_FS_CONSTANTS;

4054

}

4055

4056

static void

4057

llvmpipe_set_shader_buffers(struct pipe_context *pipe,

4058

enum pipe_shader_type shader, unsigned start_slot,

4059

unsigned count, const struct pipe_shader_buffer *buffers,

4060

unsigned writable_bitmask)

4061

{

4062

struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);

4063

unsigned i, idx;

4064

for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {

4065

const struct pipe_shader_buffer *buffer = buffers ? &buffers[idx] : NULL;

4066

4067

util_copy_shader_buffer(&llvmpipe->ssbos[shader][i], buffer);

4068

4069

if (buffer && buffer->buffer) {

4070

boolean read_only = !(writable_bitmask & (1 << idx));

4071

llvmpipe_flush_resource(pipe, buffer->buffer, 0, read_only, false,

4072

false, "buffer");

4073

}

4074

4075

if (shader == PIPE_SHADER_VERTEX ||

4076

shader == PIPE_SHADER_GEOMETRY ||

4077

shader == PIPE_SHADER_TESS_CTRL ||

4078

shader == PIPE_SHADER_TESS_EVAL) {

4079

const unsigned size = buffer ? buffer->buffer_size : 0;

4080

const ubyte *data = NULL;

4081

if (buffer && buffer->buffer)

4082

data = (ubyte *) llvmpipe_resource_data(buffer->buffer);

4083

if (data)

4084

data += buffer->buffer_offset;

4085

draw_set_mapped_shader_buffer(llvmpipe->draw, shader,

4086

i, data, size);

4087

} else if (shader == PIPE_SHADER_COMPUTE) {

4088

llvmpipe->cs_dirty |= LP_CSNEW_SSBOS;

4089

} else if (shader == PIPE_SHADER_FRAGMENT) {

4090

llvmpipe->fs_ssbo_write_mask &= ~(((1 << count) - 1) << start_slot);

4091

llvmpipe->fs_ssbo_write_mask |= writable_bitmask << start_slot;

4092

llvmpipe->dirty |= LP_NEW_FS_SSBOS;

4093

}

4094

}

4095

}

4096

4097

static void

4098

llvmpipe_set_shader_images(struct pipe_context *pipe,

4099

enum pipe_shader_type shader, unsigned start_slot,

4100

unsigned count, unsigned unbind_num_trailing_slots,

4101

const struct pipe_image_view *images)

4102

{

4103

struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);

4104

unsigned i, idx;

4105

4106

draw_flush(llvmpipe->draw);

4107

for (i = start_slot, idx = 0; i < start_slot + count; i++, idx++) {

4108

const struct pipe_image_view *image = images ? &images[idx] : NULL;

4109

4110

util_copy_image_view(&llvmpipe->images[shader][i], image);

4111

4112

if (image && image->resource) {

4113

bool read_only = !(image->access & PIPE_IMAGE_ACCESS_WRITE);

4114

llvmpipe_flush_resource(pipe, image->resource, 0, read_only, false,

4115

false, "image");

4116

}

4117

}

4118

4119

llvmpipe->num_images[shader] = start_slot + count;

4120

if (shader == PIPE_SHADER_VERTEX ||

4121

shader == PIPE_SHADER_GEOMETRY ||

4122

shader == PIPE_SHADER_TESS_CTRL ||

4123

shader == PIPE_SHADER_TESS_EVAL) {

4124

draw_set_images(llvmpipe->draw,

4125

shader,

4126

llvmpipe->images[shader],

4127

start_slot + count);

4128

} else if (shader == PIPE_SHADER_COMPUTE)

4129

llvmpipe->cs_dirty |= LP_CSNEW_IMAGES;

4130

else

4131

llvmpipe->dirty |= LP_NEW_FS_IMAGES;

4132

4133

if (unbind_num_trailing_slots) {

4134

llvmpipe_set_shader_images(pipe, shader, start_slot + count,

4135

unbind_num_trailing_slots, 0, NULL);

4136

}

4137

}

4138

4139

/**

4140

* Return the blend factor equivalent to a destination alpha of one.

4141

4142

static inline unsigned

4143

force_dst_alpha_one(unsigned factor, boolean clamped_zero)

4144

{

4145

switch(factor) {

4146

case PIPE_BLENDFACTOR_DST_ALPHA:

4147

return PIPE_BLENDFACTOR_ONE;

4148

case PIPE_BLENDFACTOR_INV_DST_ALPHA:

4149

return PIPE_BLENDFACTOR_ZERO;

4150

case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:

4151

if (clamped_zero)

4152

return PIPE_BLENDFACTOR_ZERO;

4153

else

4154

return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;

4155

}

4156

4157

return factor;

4158

}

4159

4160

4161

/**

4162

* We need to generate several variants of the fragment pipeline to match

4163

* all the combinations of the contributing state atoms.

4164

4165

* TODO: there is actually no reason to tie this to context state -- the

4166

* generated code could be cached globally in the screen.

4167

4168

static struct lp_fragment_shader_variant_key *

4169

make_variant_key(struct llvmpipe_context *lp,

4170

struct lp_fragment_shader *shader,

4171

char *store)

4172

{

4173

unsigned i;

4174

struct lp_fragment_shader_variant_key *key;

4175

4176

key = (struct lp_fragment_shader_variant_key *)store;

4177

4178

memset(key, 0, sizeof(*key));

4179

4180

if (lp->framebuffer.zsbuf) {

4181

enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;

4182

const struct util_format_description *zsbuf_desc =

4183

util_format_description(zsbuf_format);

4184

4185

if (lp->depth_stencil->depth_enabled &&

4186

util_format_has_depth(zsbuf_desc)) {

4187

key->zsbuf_format = zsbuf_format;

4188

key->depth.enabled = lp->depth_stencil->depth_enabled;

4189

key->depth.writemask = lp->depth_stencil->depth_writemask;

4190

key->depth.func = lp->depth_stencil->depth_func;

4191

}

4192

if (lp->depth_stencil->stencil[0].enabled &&

4193

util_format_has_stencil(zsbuf_desc)) {

4194

key->zsbuf_format = zsbuf_format;

4195

memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);

4196

}

4197

if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {

4198

key->resource_1d = TRUE;

4199

}

4200

key->zsbuf_nr_samples = util_res_sample_count(lp->framebuffer.zsbuf->texture);

4201

}

4202

4203

4204

* Propagate the depth clamp setting from the rasterizer state.

4205

4206

key->depth_clamp = lp->rasterizer->depth_clamp;

4207

4208

/* alpha test only applies if render buffer 0 is non-integer (or does not exist) */

4209

if (!lp->framebuffer.nr_cbufs ||

4210

!lp->framebuffer.cbufs[0] ||

4211

!util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {

4212

key->alpha.enabled = lp->depth_stencil->alpha_enabled;

4213

}

4214

if(key->alpha.enabled)

4215

key->alpha.func = lp->depth_stencil->alpha_func;

4216

/* alpha.ref_value is passed in jit_context */

4217

4218

key->flatshade = lp->rasterizer->flatshade;

4219

key->multisample = lp->rasterizer->multisample;

4220

key->no_ms_sample_mask_out = lp->rasterizer->no_ms_sample_mask_out;

4221

if (lp->active_occlusion_queries && !lp->queries_disabled) {

4222

key->occlusion_count = TRUE;

4223

}

4224

4225

memcpy(&key->blend, lp->blend, sizeof key->blend);

4226

4227

key->coverage_samples = 1;

4228

key->min_samples = 1;

4229

if (key->multisample) {

4230

key->coverage_samples = util_framebuffer_get_num_samples(&lp->framebuffer);

4231

key->min_samples = lp->min_samples == 1 ? 1 : key->coverage_samples;

4232

}

4233

key->nr_cbufs = lp->framebuffer.nr_cbufs;

4234

4235

if (!key->blend.independent_blend_enable) {

4236

/* we always need independent blend otherwise the fixups below won't work */

4237

for (i = 1; i < key->nr_cbufs; i++) {

4238

memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));

4239

}

4240

key->blend.independent_blend_enable = 1;

4241

}

4242

4243

for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {

4244

struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];

4245

4246

if (lp->framebuffer.cbufs[i]) {

4247

enum pipe_format format = lp->framebuffer.cbufs[i]->format;

4248

const struct util_format_description *format_desc;

4249

4250

key->cbuf_format[i] = format;

4251

key->cbuf_nr_samples[i] = util_res_sample_count(lp->framebuffer.cbufs[i]->texture);

4252

4253

4254

* Figure out if this is a 1d resource. Note that OpenGL allows crazy

4255

* mixing of 2d textures with height 1 and 1d textures, so make sure

4256

* we pick 1d if any cbuf or zsbuf is 1d.

4257

4258

if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {

4259

key->resource_1d = TRUE;

4260

}

4261

4262

format_desc = util_format_description(format);

4263

assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||

4264

format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);

4265

4266

4267

* Mask out color channels not present in the color buffer.

4268

4269

blend_rt->colormask &= util_format_colormask(format_desc);

4270

4271

4272

* Disable blend for integer formats.

4273

4274

if (util_format_is_pure_integer(format)) {

4275

blend_rt->blend_enable = 0;

4276

}

4277

4278

4279

* Our swizzled render tiles always have an alpha channel, but the

4280

* linear render target format often does not, so force here the dst

4281

* alpha to be one.

4282

4283

* This is not a mere optimization. Wrong results will be produced if

4284

* the dst alpha is used, the dst format does not have alpha, and the

4285

* previous rendering was not flushed from the swizzled to linear

4286

* buffer. For example, NonPowTwo DCT.

4287

4288

* TODO: This should be generalized to all channels for better

4289

* performance, but only alpha causes correctness issues.

4290

4291

* Also, force rgb/alpha func/factors match, to make AoS blending

4292

* easier.

4293

4294

if (format_desc->swizzle[3] > PIPE_SWIZZLE_W ||

4295

format_desc->swizzle[3] == format_desc->swizzle[0]) {

4296

/* Doesn't cover mixed snorm/unorm but can't render to them anyway */

4297

boolean clamped_zero = !util_format_is_float(format) &&

4298

!util_format_is_snorm(format);

4299

blend_rt->rgb_src_factor =

4300

force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);

4301

blend_rt->rgb_dst_factor =

4302

force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);

4303

blend_rt->alpha_func = blend_rt->rgb_func;

4304

blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;

4305

blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;

4306

}

4307

}

4308

else {

4309

/* no color buffer for this fragment output */

4310

key->cbuf_format[i] = PIPE_FORMAT_NONE;

4311

key->cbuf_nr_samples[i] = 0;

4312

blend_rt->colormask = 0x0;

4313

blend_rt->blend_enable = 0;

4314

}

4315

}

4316

4317

/* This value will be the same for all the variants of a given shader:

4318

4319

key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;

4320

4321

if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1)

4322

key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;

4323

4324

struct lp_sampler_static_state *fs_sampler;

4325

4326

fs_sampler = lp_fs_variant_key_samplers(key);

4327

4328

memset(fs_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *fs_sampler);

4329

4330

for(i = 0; i < key->nr_samplers; ++i) {

4331

if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {

4332

lp_sampler_static_sampler_state(&fs_sampler[i].sampler_state,

4333

lp->samplers[PIPE_SHADER_FRAGMENT][i]);

4334

}

4335

}

4336

4337

4338

* XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes

4339

* are dx10-style? Can't really have mixed opcodes, at least not

4340

* if we want to skip the holes here (without rescanning tgsi).

4341

4342

if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {

4343

for(i = 0; i < key->nr_sampler_views; ++i) {

4344

4345

* Note sview may exceed what's representable by file_mask.

4346

* This will still work, the only downside is that not actually

4347

* used views may be included in the shader key.

4348

4349

if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1u << (i & 31))) {

4350

lp_sampler_static_texture_state(&fs_sampler[i].texture_state,

4351

lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);

4352

}

4353

}

4354

}

4355

else {

4356

key->nr_sampler_views = key->nr_samplers;

4357

for(i = 0; i < key->nr_sampler_views; ++i) {

4358

if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {

4359

lp_sampler_static_texture_state(&fs_sampler[i].texture_state,

4360

lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);

4361

}

4362

}

4363

}

4364

4365

struct lp_image_static_state *lp_image;

4366

lp_image = lp_fs_variant_key_images(key);

4367

key->nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;

4368

for (i = 0; i < key->nr_images; ++i) {

4369

if (shader->info.base.file_mask[TGSI_FILE_IMAGE] & (1 << i)) {

4370

lp_sampler_static_texture_state_image(&lp_image[i].image_state,

4371

&lp->images[PIPE_SHADER_FRAGMENT][i]);

4372

}

4373

}

4374

4375

if (shader->kind == LP_FS_KIND_AERO_MINIFICATION) {

4376

struct lp_sampler_static_state *samp0 = lp_fs_variant_key_sampler_idx(key, 0);

4377

assert(samp0);

4378

samp0->sampler_state.min_img_filter = PIPE_TEX_FILTER_NEAREST;

4379

samp0->sampler_state.mag_img_filter = PIPE_TEX_FILTER_NEAREST;

4380

}

4381

4382

return key;

4383

}

4384

4385

4386

/**

4387

* Update fragment shader state. This is called just prior to drawing

4388

* something when some fragment-related state has changed.

4389

4390

void

4391

llvmpipe_update_fs(struct llvmpipe_context *lp)

4392

{

4393

struct lp_fragment_shader *shader = lp->fs;

4394

struct lp_fragment_shader_variant_key *key;

4395

struct lp_fragment_shader_variant *variant = NULL;

4396

struct lp_fs_variant_list_item *li;

4397

char store[LP_FS_MAX_VARIANT_KEY_SIZE];

4398

4399

key = make_variant_key(lp, shader, store);

4400

4401

/* Search the variants for one which matches the key */

4402

li = first_elem(&shader->variants);

4403

while(!at_end(&shader->variants, li)) {

4404

if(memcmp(&li->base->key, key, shader->variant_key_size) == 0) {

4405

variant = li->base;

4406

break;

4407

}

4408

li = next_elem(li);

4409

}

4410

4411

if (variant) {

4412

/* Move this variant to the head of the list to implement LRU

4413

* deletion of shader's when we have too many.

4414

4415

move_to_head(&lp->fs_variants_list, &variant->list_item_global);

4416

}

4417

else {

4418

/* variant not found, create it now */

4419

int64_t t0, t1, dt;

4420

unsigned i;

4421

unsigned variants_to_cull;

4422

4423

if (LP_DEBUG & DEBUG_FS) {

4424

debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",

4425

lp->nr_fs_variants,

4426

lp->nr_fs_instrs,

4427

lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);

4428

}

4429

4430

/* First, check if we've exceeded the max number of shader variants.

4431

* If so, free 6.25% of them (the least recently used ones).

4432

4433

variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 16 : 0;

4434

4435

if (variants_to_cull ||

4436

lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {

4437

if (gallivm_debug & GALLIVM_DEBUG_PERF) {

4438

debug_printf("Evicting FS: %u fs variants,\t%u total variants,"

4439

"\t%u instrs,\t%u instrs/variant\n",

4440

shader->variants_cached,

4441

lp->nr_fs_variants, lp->nr_fs_instrs,

4442

lp->nr_fs_instrs / lp->nr_fs_variants);

4443

}

4444

4445

4446

* We need to re-check lp->nr_fs_variants because an arbitrarliy large

4447

* number of shader variants (potentially all of them) could be

4448

* pending for destruction on flush.

4449

4450

4451

for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {

4452

struct lp_fs_variant_list_item *item;

4453

if (is_empty_list(&lp->fs_variants_list)) {

4454

break;

4455

}

4456

item = last_elem(&lp->fs_variants_list);

4457

assert(item);

4458

assert(item->base);

4459

llvmpipe_remove_shader_variant(lp, item->base);

4460

struct lp_fragment_shader_variant *variant = item->base;

4461

lp_fs_variant_reference(lp, &variant, NULL);

4462

}

4463

}

4464

4465

4466

* Generate the new variant.

4467

4468

t0 = os_time_get();

4469

variant = generate_variant(lp, shader, key);

4470

t1 = os_time_get();

4471

dt = t1 - t0;

4472

LP_COUNT_ADD(llvm_compile_time, dt);

4473

LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */

4474

4475

/* Put the new variant into the list */

4476

if (variant) {

4477

insert_at_head(&shader->variants, &variant->list_item_local);

4478

insert_at_head(&lp->fs_variants_list, &variant->list_item_global);

4479

lp->nr_fs_variants++;

4480

lp->nr_fs_instrs += variant->nr_instrs;

4481

shader->variants_cached++;

4482

}

4483

}

4484

4485

/* Bind this variant */

4486

lp_setup_set_fs_variant(lp->setup, variant);

4487

}

4488

4489

4490

4491

4492

4493

void

4494

llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)

4495

{

4496

llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;

4497

llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state;

4498

llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;

4499

4500

llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;

4501

4502

llvmpipe->pipe.set_shader_buffers = llvmpipe_set_shader_buffers;

4503

llvmpipe->pipe.set_shader_images = llvmpipe_set_shader_images;

4504

}

4505

4506

Older »