[virglrenderer-devel] [PATCH 3/4] shader: add basic shader_storage_buffer_object parsing. (v2)
Tomeu Vizoso
tomeu.vizoso at collabora.com
Wed Jul 18 11:10:50 UTC 2018
On 07/18/2018 04:26 AM, Dave Airlie wrote:
> From: Dave Airlie <airlied at redhat.com>
>
> This adds the basic shader parsing for the SSBO extension,
>
> v2: drop qualifier, cleanup unused var, add indirect support,
> use a bitmask to track declared ssbos.
> ---
> src/vrend_shader.c | 331 ++++++++++++++++++++++++++++++++++++++++++++++++-----
> src/vrend_shader.h | 1 +
> 2 files changed, 305 insertions(+), 27 deletions(-)
>
> diff --git a/src/vrend_shader.c b/src/vrend_shader.c
> index cbbfbbc..b20d681 100644
> --- a/src/vrend_shader.c
> +++ b/src/vrend_shader.c
> @@ -108,6 +108,28 @@ struct vrend_io_range {
> bool used;
> };
>
> +enum vrend_type_qualifier {
Apparently this enum doesn't need to be moved around?
> + 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 dump_ctx {
> struct tgsi_iterate_context iter;
> struct vrend_shader_cfg *cfg;
> @@ -136,7 +158,12 @@ struct dump_ctx {
>
> struct vrend_shader_sampler samplers[32];
> uint32_t samplers_used;
> - bool sviews_used;
> +
> + uint32_t ssbo_used_mask;
> + uint32_t ssbo_atomic_mask;
> + uint32_t ssbo_array_base;
> + uint32_t ssbo_atomic_array_base;
> +
> struct vrend_sampler_array *sampler_arrays;
> uint32_t num_sampler_arrays;
> int last_sampler_array_idx;
> @@ -206,28 +233,6 @@ 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;
> @@ -1040,8 +1045,22 @@ iter_declaration(struct tgsi_iterate_context *iter,
> } else {
> ctx->last_sampler_array_idx = add_sampler_array(ctx, decl->Range.First, decl->Range.Last + 1, decl->SamplerView.Resource, decl->SamplerView.ReturnTypeX);
> }
> - } else
> - ctx->sviews_used = true;
Was this removal intentional?
> + }
> + break;
> + case TGSI_FILE_BUFFER:
> + if (decl->Range.First >= 32) {
> + fprintf(stderr, "Buffer view exceeded, max is 32\n");
> + return FALSE;
> + }
> + ctx->ssbo_used_mask |= (1 << decl->Range.First);
> + if (decl->Declaration.Atomic) {
> + if (decl->Range.First < ctx->ssbo_atomic_array_base)
> + ctx->ssbo_atomic_array_base = decl->Range.First;
> + ctx->ssbo_atomic_mask |= (1 << decl->Range.First);
> + } else {
> + if (decl->Range.First < ctx->ssbo_array_base)
> + ctx->ssbo_array_base = decl->Range.First;
> + }
> break;
> case TGSI_FILE_CONSTANT:
> if (decl->Declaration.Dimension && decl->Dim.Index2D != 0) {
> @@ -2123,6 +2142,173 @@ 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,
> + 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 ... */
conversion is used unconditionally, so this comment doesn't make sense here.
> + 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]);
s/floatBitsToUint/conversion as we already have it?
> + 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';
> + if (ctx->ssbo_atomic_mask & (1 << src->Register.Index)) {
> + /* Emulate atomicCounter with atomicOr. */
> + strcpy(atomic_op, "atomicOr");
> + strcpy(atomic_src, ", uint(0)");
> + }
> +
> + 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);
Guess we don't want to emit an uninitialized buffer?
> + 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);
Same comment as before.
Cheers,
Tomeu
> + return 0;
> +}
> +
> static int
> get_destination_info(struct dump_ctx *ctx,
> const struct tgsi_full_instruction *inst,
> @@ -2278,8 +2464,19 @@ 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 (ctx->info.indirect_files & (1 << TGSI_FILE_BUFFER)) {
> + bool atomic_ssbo = ctx->ssbo_atomic_mask & (1 << dst_reg->Register.Index);
> + const char *atomic_str = atomic_ssbo ? "atomic" : "";
> + int base = atomic_ssbo ? ctx->ssbo_atomic_array_base : ctx->ssbo_array_base;
> + if (dst_reg->Register.Indirect) {
> + snprintf(dsts[i], 255, "%sssboarr%s[addr%d+%d].%sssbocontents%d", cname, atomic_str, dst_reg->Indirect.Index, dst_reg->Register.Index - base, cname, base);
> + } else
> + snprintf(dsts[i], 255, "%sssboarr%s[%d].%sssbocontents%d", cname, atomic_str, dst_reg->Register.Index - base, cname, base);
> + } 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 +2732,21 @@ 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)) {
> + bool atomic_ssbo = ctx->ssbo_atomic_mask & (1 << src->Register.Index);
> + const char *atomic_str = atomic_ssbo ? "atomic" : "";
> + int base = atomic_ssbo ? ctx->ssbo_atomic_array_base : ctx->ssbo_array_base;
> + if (src->Register.Indirect) {
> + snprintf(srcs[i], 255, "%sssboarr%s[addr%d+%d].%sssbocontents%d%s", cname, atomic_str, src->Indirect.Index, src->Register.Index - base, cname, base, swizzle);
> + } else {
> + snprintf(srcs[i], 255, "%sssboarr%s[%d].%sssbocontents%d%s", cname, atomic_str, src->Register.Index - base, cname, base, swizzle);
> + }
> + } 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 +3476,35 @@ 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, 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 +3585,8 @@ 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->ssbo_used_mask)
> + 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 +4044,11 @@ static char *emit_ios(struct dump_ctx *ctx, char *glsl_hdr)
> STRCAT_WITH_RET(glsl_hdr, buf);
> }
>
> + if (ctx->ssbo_used_mask) {
> + 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 +4112,26 @@ static char *emit_ios(struct dump_ctx *ctx, char *glsl_hdr)
> return NULL;
> }
> }
> +
> + if (ctx->info.indirect_files & (1 << TGSI_FILE_BUFFER)) {
> + uint32_t mask = ctx->ssbo_used_mask;
> + while (mask) {
> + int start, count, i;
> + u_bit_scan_consecutive_range(&mask, &start, &count);
> + const char *atomic = (ctx->ssbo_atomic_mask & (1 << start)) ? "atomic" : "";
> + snprintf(buf, 255, "layout (binding = %d, std430) buffer %sssbo%d { uint %sssbocontents%d[]; } %sssboarr%s[%d];\n", start, sname, start, sname, start, sname, atomic, count);
> + STRCAT_WITH_RET(glsl_hdr, buf);
> + }
> + } else {
> + uint32_t mask = ctx->ssbo_used_mask;
> + while (mask) {
> + uint32_t id = u_bit_scan(&mask);
> + 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");
> @@ -3944,6 +4212,8 @@ char *vrend_convert_shader(struct vrend_shader_cfg *cfg,
> ctx.num_sampler_arrays = 0;
> ctx.sampler_arrays = NULL;
> ctx.last_sampler_array_idx = -1;
> + ctx.ssbo_array_base = 0xffffffff;
> + ctx.ssbo_atomic_array_base = 0xffffffff;
> ctx.has_sample_input = false;
> tgsi_scan_shader(tokens, &ctx.info);
> /* if we are in core profile mode we should use GLSL 1.40 */
> @@ -3961,6 +4231,10 @@ char *vrend_convert_shader(struct vrend_shader_cfg *cfg,
> if (ctx.info.dimension_indirect_files & (1 << TGSI_FILE_CONSTANT))
> require_glsl_ver(&ctx, 150);
>
> + if (ctx.info.indirect_files & (1 << TGSI_FILE_BUFFER)) {
> + require_glsl_ver(&ctx, 150);
> + ctx.shader_req_bits |= SHADER_REQ_GPU_SHADER5;
> + }
> if (ctx.info.indirect_files & (1 << TGSI_FILE_SAMPLER))
> ctx.shader_req_bits |= SHADER_REQ_GPU_SHADER5;
>
> @@ -4012,6 +4286,9 @@ 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));
> +
> + sinfo->ssbo_used_mask = ctx.ssbo_used_mask;
> +
> 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;
>
More information about the virglrenderer-devel
mailing list