[virglrenderer-devel] [PATCH 2/3] shader: add basic shader_storage_buffer_object parsing.
Dave Airlie
airlied at gmail.com
Tue Jul 17 07:48:09 UTC 2018
From: Dave Airlie <airlied at redhat.com>
This adds the basic shader parsing for the SSBO extension,
TODO:
Add indirect handling.
---
src/vrend_shader.c | 304 +++++++++++++++++++++++++++++++++++++++++----
src/vrend_shader.h | 1 +
2 files changed, 281 insertions(+), 24 deletions(-)
diff --git a/src/vrend_shader.c b/src/vrend_shader.c
index cbbfbbc..ae3f028 100644
--- a/src/vrend_shader.c
+++ b/src/vrend_shader.c
@@ -108,6 +108,34 @@ struct vrend_io_range {
bool used;
};
+enum vrend_type_qualifier {
+ TYPE_CONVERSION_NONE = 0,
+ FLOAT = 1,
+ VEC2 = 2,
+ VEC3 = 3,
+ VEC4 = 4,
+ INT = 5,
+ IVEC2 = 6,
+ IVEC3 = 7,
+ IVEC4 = 8,
+ UINT = 9,
+ UVEC2 = 10,
+ UVEC3 = 11,
+ UVEC4 = 12,
+ FLOAT_BITS_TO_UINT = 13,
+ UINT_BITS_TO_FLOAT = 14,
+ FLOAT_BITS_TO_INT = 15,
+ INT_BITS_TO_FLOAT = 16,
+ DOUBLE = 17,
+ DVEC2 = 18,
+};
+
+struct vrend_ssbo {
+ uint32_t id;
+ unsigned atomic;
+ enum vrend_type_qualifier type;
+};
+
struct dump_ctx {
struct tgsi_iterate_context iter;
struct vrend_shader_cfg *cfg;
@@ -137,6 +165,10 @@ struct dump_ctx {
struct vrend_shader_sampler samplers[32];
uint32_t samplers_used;
bool sviews_used;
+
+ uint32_t num_ssbo;
+ struct vrend_ssbo ssbos[32];
+
struct vrend_sampler_array *sampler_arrays;
uint32_t num_sampler_arrays;
int last_sampler_array_idx;
@@ -206,28 +238,6 @@ static const struct vrend_shader_table shader_req_table[] = {
{ SHADER_REQ_FP64, "GL_ARB_gpu_shader_fp64" },
};
-enum vrend_type_qualifier {
- TYPE_CONVERSION_NONE = 0,
- FLOAT = 1,
- VEC2 = 2,
- VEC3 = 3,
- VEC4 = 4,
- INT = 5,
- IVEC2 = 6,
- IVEC3 = 7,
- IVEC4 = 8,
- UINT = 9,
- UVEC2 = 10,
- UVEC3 = 11,
- UVEC4 = 12,
- FLOAT_BITS_TO_UINT = 13,
- UINT_BITS_TO_FLOAT = 14,
- FLOAT_BITS_TO_INT = 15,
- INT_BITS_TO_FLOAT = 16,
- DOUBLE = 17,
- DVEC2 = 18,
-};
-
struct dest_info {
enum vrend_type_qualifier dtypeprefix;
enum vrend_type_qualifier dstconv;
@@ -1043,6 +1053,15 @@ iter_declaration(struct tgsi_iterate_context *iter,
} else
ctx->sviews_used = true;
break;
+ case TGSI_FILE_BUFFER:
+ if (ctx->num_ssbo >= ARRAY_SIZE(ctx->ssbos)) {
+ fprintf(stderr, "Buffer view exceeded, max is %lu\n", ARRAY_SIZE(ctx->ssbos));
+ return FALSE;
+ }
+ ctx->ssbos[ctx->num_ssbo].id = decl->Range.First;
+ ctx->ssbos[ctx->num_ssbo].atomic = decl->Declaration.Atomic;
+ ctx->num_ssbo++;
+ break;
case TGSI_FILE_CONSTANT:
if (decl->Declaration.Dimension && decl->Dim.Index2D != 0) {
if (ctx->num_ubo >= ARRAY_SIZE(ctx->ubo_idx)) {
@@ -2123,6 +2142,177 @@ create_swizzled_clipdist(struct dump_ctx *ctx,
snprintf(result, 255, "%s(vec4(%s,%s,%s,%s))", stypeprefix, clipdistvec[0], clipdistvec[1], clipdistvec[2], clipdistvec[3]);
}
+static int
+translate_store(struct dump_ctx *ctx,
+ struct tgsi_full_instruction *inst,
+ struct source_info *sinfo,
+ char srcs[4][255],
+ char dsts[3][255])
+{
+ const struct tgsi_full_dst_register *dst = &inst->Dst[0];
+ char buf[512];
+
+ if (dst->Register.File == TGSI_FILE_BUFFER) {
+ /* We don't want to cast uvec4's ... */
+ const char *conversion = get_string(FLOAT_BITS_TO_UINT);
+ if (inst->Dst[0].Register.WriteMask & 0x1) {
+ snprintf(buf, 255, "%s[uint(floatBitsToUint(%s))>>2] = %s(%s).x;\n", dsts[0], srcs[0], conversion, srcs[1]);
+ EMIT_BUF_WITH_RET(ctx, buf);
+ }
+ if (inst->Dst[0].Register.WriteMask & 0x2) {
+ snprintf(buf, 255, "%s[(uint(floatBitsToUint(%s))>>2)+1u] = %s(%s).y;\n", dsts[0], srcs[0], conversion, srcs[1]);
+ EMIT_BUF_WITH_RET(ctx, buf);
+ }
+ if (inst->Dst[0].Register.WriteMask & 0x4) {
+ snprintf(buf, 255, "%s[(uint(floatBitsToUint(%s))>>2)+2u] = %s(%s).z;\n", dsts[0], srcs[0], conversion, srcs[1]);
+ EMIT_BUF_WITH_RET(ctx, buf);
+ }
+ if (inst->Dst[0].Register.WriteMask & 0x8) {
+ snprintf(buf, 255, "%s[(uint(floatBitsToUint(%s))>>2)+3u] = %s(%s).w;\n", dsts[0], srcs[0], conversion, srcs[1]);
+ EMIT_BUF_WITH_RET(ctx, buf);
+ }
+ }
+ return 0;
+}
+
+static int
+translate_load(struct dump_ctx *ctx,
+ struct tgsi_full_instruction *inst,
+ char srcs[4][255],
+ char dsts[3][255])
+{
+ char buf[512];
+ const struct tgsi_full_src_register *src = &inst->Src[0];
+
+ if (src->Register.File == TGSI_FILE_BUFFER) {
+ char mydst[255], atomic_op[9], atomic_src[10];
+ strcpy(mydst, dsts[0]);
+ char *wmp = strchr(mydst, '.');
+ if (wmp)
+ wmp[0] = 0;
+ snprintf(buf, 255, "ssbo_addr_temp = uint(floatBitsToUint(%s)) >> 2;\n", srcs[1]);
+ EMIT_BUF_WITH_RET(ctx, buf);
+
+ atomic_op[0] = atomic_src[0] = '\0';
+ for (uint32_t i = 0; i < ctx->num_ssbo; i++) {
+ /* Emulate atomicCounter with atomicOr. */
+ if ((ctx->ssbos[i].id == (uint32_t)src->Register.Index) && ctx->ssbos[i].atomic) {
+ strcpy(atomic_op, "atomicOr");
+ strcpy(atomic_src, ", uint(0)");
+ break;
+ }
+ }
+
+ if (inst->Dst[0].Register.WriteMask & 0x1) {
+ snprintf(buf, 255, "%s.x = (uintBitsToFloat(%s(%s[ssbo_addr_temp]%s)));\n", mydst, atomic_op, srcs[0], atomic_src);
+ EMIT_BUF_WITH_RET(ctx, buf);
+ }
+ if (inst->Dst[0].Register.WriteMask & 0x2) {
+ snprintf(buf, 255, "%s.y = (uintBitsToFloat(%s(%s[ssbo_addr_temp + 1u]%s)));\n", mydst, atomic_op, srcs[0], atomic_src);
+ EMIT_BUF_WITH_RET(ctx, buf);
+ }
+ if (inst->Dst[0].Register.WriteMask & 0x4) {
+ snprintf(buf, 255, "%s.z = (uintBitsToFloat(%s(%s[ssbo_addr_temp + 2u]%s)));\n", mydst, atomic_op, srcs[0], atomic_src);
+ EMIT_BUF_WITH_RET(ctx, buf);
+ }
+ if (inst->Dst[0].Register.WriteMask & 0x8) {
+ snprintf(buf, 255, "%s.w = (uintBitsToFloat(%s(%s[ssbo_addr_temp + 3u]%s)));\n", mydst, atomic_op, srcs[0], atomic_src);
+ EMIT_BUF_WITH_RET(ctx, buf);
+ }
+ }
+ return 0;
+}
+
+static const char *get_atomic_opname(int tgsi_opcode, bool *is_cas)
+{
+ const char *opname;
+ *is_cas = false;
+ switch (tgsi_opcode) {
+ case TGSI_OPCODE_ATOMUADD:
+ opname = "Add";
+ break;
+ case TGSI_OPCODE_ATOMXCHG:
+ opname = "Exchange";
+ break;
+ case TGSI_OPCODE_ATOMCAS:
+ opname = "CompSwap";
+ *is_cas = true;
+ break;
+ case TGSI_OPCODE_ATOMAND:
+ opname = "And";
+ break;
+ case TGSI_OPCODE_ATOMOR:
+ opname = "Or";
+ break;
+ case TGSI_OPCODE_ATOMXOR:
+ opname = "Xor";
+ break;
+ case TGSI_OPCODE_ATOMUMIN:
+ opname = "Min";
+ break;
+ case TGSI_OPCODE_ATOMUMAX:
+ opname = "Max";
+ break;
+ case TGSI_OPCODE_ATOMIMIN:
+ opname = "Min";
+ break;
+ case TGSI_OPCODE_ATOMIMAX:
+ opname = "Max";
+ break;
+ default:
+ fprintf(stderr, "illegal atomic opcode");
+ return NULL;
+ }
+ return opname;
+}
+
+static int
+translate_resq(struct dump_ctx *ctx, struct tgsi_full_instruction *inst,
+ char srcs[4][255], char dsts[3][255])
+{
+ char buf[512];
+ const struct tgsi_full_src_register *src = &inst->Src[0];
+
+ if (src->Register.File == TGSI_FILE_BUFFER) {
+ snprintf(buf, 255, "%s = %s(int(%s.length()) << 2);\n", dsts[0], get_string(INT_BITS_TO_FLOAT), srcs[0]);
+ }
+
+ EMIT_BUF_WITH_RET(ctx, buf);
+ return 0;
+}
+
+static int
+translate_atomic(struct dump_ctx *ctx,
+ struct tgsi_full_instruction *inst,
+ char srcs[4][255],
+ char dsts[3][255])
+{
+ char buf[512];
+ const struct tgsi_full_src_register *src = &inst->Src[0];
+ const char *opname;
+ enum vrend_type_qualifier stypeprefix;
+ enum vrend_type_qualifier dtypeprefix;
+ enum vrend_type_qualifier stypecast;
+ bool is_cas;
+ char cas_str[64] = {};
+
+ stypeprefix = FLOAT_BITS_TO_UINT;
+ dtypeprefix = UINT_BITS_TO_FLOAT;
+ stypecast = INT;
+
+ opname = get_atomic_opname(inst->Instruction.Opcode, &is_cas);
+ if (!opname)
+ return -1;
+
+ if (is_cas)
+ snprintf(cas_str, 64, ", %s(%s(%s))", get_string(stypecast), get_string(stypeprefix), srcs[3]);
+
+ if (src->Register.File == TGSI_FILE_BUFFER)
+ snprintf(buf, 512, "%s = %s(atomic%s(%s[int(floatBitsToInt(%s)) >> 2], uint(%s(%s).x)%s));\n", dsts[0], get_string(dtypeprefix), opname, srcs[0], srcs[1], get_string(stypeprefix), srcs[2], cas_str);
+ EMIT_BUF_WITH_RET(ctx, buf);
+ return 0;
+}
+
static int
get_destination_info(struct dump_ctx *ctx,
const struct tgsi_full_instruction *inst,
@@ -2278,8 +2468,14 @@ get_destination_info(struct dump_ctx *ctx,
snprintf(dsts[i], 255, "temp%d[addr0 + %d]%s", range->first, dst_reg->Register.Index - range->first, writemask);
} else
snprintf(dsts[i], 255, "temp%d[%d]%s", range->first, dst_reg->Register.Index - range->first, writemask);
- }
- else if (dst_reg->Register.File == TGSI_FILE_ADDRESS) {
+ } else if (dst_reg->Register.File == TGSI_FILE_BUFFER) {
+ const char *cname = tgsi_proc_to_prefix(ctx->prog_type);
+ if (dst_reg->Register.Indirect) {
+ assert(dst_reg->Indirect.File == TGSI_FILE_ADDRESS);
+ snprintf(dsts[i], 255, "%sssbocontents[%d]", cname, dst_reg->Register.Index);
+ } else
+ snprintf(dsts[i], 255, "%sssbocontents%d", cname, dst_reg->Register.Index);
+ } else if (dst_reg->Register.File == TGSI_FILE_ADDRESS) {
snprintf(dsts[i], 255, "addr%d", dst_reg->Register.Index);
}
@@ -2535,6 +2731,14 @@ get_source_info(struct dump_ctx *ctx,
snprintf(srcs[i], 255, "%ssamp%d%s", cname, src->Register.Index, swizzle);
}
sinfo->sreg_index = src->Register.Index;
+ } else if (src->Register.File == TGSI_FILE_BUFFER) {
+ const char *cname = tgsi_proc_to_prefix(ctx->prog_type);
+ if (ctx->info.indirect_files & (1 << TGSI_FILE_BUFFER)) {
+ fprintf(stderr, "TODO\n");
+ } else {
+ snprintf(srcs[i], 255, "%sssbocontents%d%s", cname, src->Register.Index, swizzle);
+ }
+ sinfo->sreg_index = src->Register.Index;
} else if (src->Register.File == TGSI_FILE_IMMEDIATE) {
if (src->Register.Index >= (int)ARRAY_SIZE(ctx->imm)) {
fprintf(stderr, "Immediate exceeded, max is %lu\n", ARRAY_SIZE(ctx->imm));
@@ -3264,6 +3468,35 @@ iter_instruction(struct tgsi_iterate_context *iter,
snprintf(buf, 255, "barrier();\n");
EMIT_BUF_WITH_RET(ctx, buf);
break;
+ case TGSI_OPCODE_STORE:
+ ret = translate_store(ctx, inst, &sinfo, srcs, dsts);
+ if (ret)
+ return FALSE;
+ break;
+ case TGSI_OPCODE_LOAD:
+ ret = translate_load(ctx, inst, srcs, dsts);
+ if (ret)
+ return FALSE;
+ break;
+ case TGSI_OPCODE_ATOMUADD:
+ case TGSI_OPCODE_ATOMXCHG:
+ case TGSI_OPCODE_ATOMCAS:
+ case TGSI_OPCODE_ATOMAND:
+ case TGSI_OPCODE_ATOMOR:
+ case TGSI_OPCODE_ATOMXOR:
+ case TGSI_OPCODE_ATOMUMIN:
+ case TGSI_OPCODE_ATOMUMAX:
+ case TGSI_OPCODE_ATOMIMIN:
+ case TGSI_OPCODE_ATOMIMAX:
+ ret = translate_atomic(ctx, inst, srcs, dsts);
+ if (ret)
+ return FALSE;
+ break;
+ case TGSI_OPCODE_RESQ:
+ ret = translate_resq(ctx, inst, srcs, dsts);
+ if (ret)
+ return FALSE;
+ break;
default:
fprintf(stderr,"failed to convert opcode %d\n", inst->Instruction.Opcode);
break;
@@ -3344,6 +3577,8 @@ static char *emit_header(struct dump_ctx *ctx, char *glsl_hdr)
if (ctx->num_cull_dist_prop || ctx->key->prev_stage_num_cull_out)
STRCAT_WITH_RET(glsl_hdr, "#extension GL_ARB_cull_distance : require\n");
+ if (ctx->num_ssbo)
+ STRCAT_WITH_RET(glsl_hdr, "#extension GL_ARB_shader_storage_buffer_object : require\n");
for (uint32_t i = 0; i < ARRAY_SIZE(shader_req_table); i++) {
if (shader_req_table[i].key == SHADER_REQ_SAMPLER_RECT && ctx->glsl_ver_required >= 140)
@@ -3801,6 +4036,11 @@ static char *emit_ios(struct dump_ctx *ctx, char *glsl_hdr)
STRCAT_WITH_RET(glsl_hdr, buf);
}
+ if (ctx->num_ssbo) {
+ snprintf(buf, 255, "uint ssbo_addr_temp;\n");
+ STRCAT_WITH_RET(glsl_hdr, buf);
+ }
+
if (ctx->shader_req_bits & SHADER_REQ_FP64) {
snprintf(buf, 255, "dvec2 fp64_dst[3];\n");
STRCAT_WITH_RET(glsl_hdr, buf);
@@ -3864,6 +4104,18 @@ static char *emit_ios(struct dump_ctx *ctx, char *glsl_hdr)
return NULL;
}
}
+
+ if (ctx->info.indirect_files & (1 << TGSI_FILE_BUFFER)) {
+
+ } else {
+ for (i = 0; i < ctx->num_ssbo; i++){
+ uint32_t id = ctx->ssbos[i].id;
+ sname = tgsi_proc_to_prefix(ctx->prog_type);
+ snprintf(buf, 255, "layout (binding = %d, std430) buffer %sssbo%d { uint %sssbocontents%d[]; };\n", id, sname, id, sname, id);
+ STRCAT_WITH_RET(glsl_hdr, buf);
+ }
+ }
+
if (ctx->prog_type == TGSI_PROCESSOR_FRAGMENT &&
ctx->key->pstipple_tex == true) {
snprintf(buf, 255, "uniform sampler2D pstipple_sampler;\nfloat stip_temp;\n");
@@ -4012,6 +4264,10 @@ char *vrend_convert_shader(struct vrend_shader_cfg *cfg,
sinfo->num_consts = ctx.num_consts;
sinfo->num_ubos = ctx.num_ubo;
memcpy(sinfo->ubo_idx, ctx.ubo_idx, ctx.num_ubo * sizeof(*ctx.ubo_idx));
+
+ for (uint32_t i = 0; i < ctx.num_ssbo; i++)
+ sinfo->ssbo_used_mask |= (1 << ctx.ssbos[i].id);
+
sinfo->ubo_indirect = ctx.info.dimension_indirect_files & (1 << TGSI_FILE_CONSTANT);
if (ctx_indirect_inputs(&ctx)) {
if (ctx.generic_input_range.used)
diff --git a/src/vrend_shader.h b/src/vrend_shader.h
index db6f135..a19da20 100644
--- a/src/vrend_shader.h
+++ b/src/vrend_shader.h
@@ -46,6 +46,7 @@ struct vrend_sampler_array {
struct vrend_shader_info {
uint32_t samplers_used_mask;
+ uint32_t ssbo_used_mask;
int num_consts;
int num_inputs;
int num_interps;
--
2.17.1
More information about the virglrenderer-devel
mailing list