[PATCH v5 0/1] MR74: vkd3d-shader/hlsl: Support dot() for SM1.
-- v5: vkd3d-shader/hlsl: Support dot() for SM1. https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/74
From: Nikolay Sivov <nsivov(a)codeweavers.com> --- libs/vkd3d-shader/hlsl.c | 1 + libs/vkd3d-shader/hlsl.h | 1 + libs/vkd3d-shader/hlsl_codegen.c | 57 ++++++++++++++++++++++++++++++++ libs/vkd3d-shader/hlsl_sm1.c | 57 +++++++++++++++++++++++++++++++- 4 files changed, 115 insertions(+), 1 deletion(-) diff --git a/libs/vkd3d-shader/hlsl.c b/libs/vkd3d-shader/hlsl.c index 256e466a..6b8e1b10 100644 --- a/libs/vkd3d-shader/hlsl.c +++ b/libs/vkd3d-shader/hlsl.c @@ -1768,6 +1768,7 @@ const char *debug_hlsl_expr_op(enum hlsl_ir_expr_op op) [HLSL_OP2_NEQUAL] = "!=", [HLSL_OP2_RSHIFT] = ">>", + [HLSL_OP3_DP2ADD] = "dp2add", [HLSL_OP3_LERP] = "lerp", }; diff --git a/libs/vkd3d-shader/hlsl.h b/libs/vkd3d-shader/hlsl.h index bb63f827..2d0beb6c 100644 --- a/libs/vkd3d-shader/hlsl.h +++ b/libs/vkd3d-shader/hlsl.h @@ -497,6 +497,7 @@ enum hlsl_ir_expr_op HLSL_OP2_NEQUAL, HLSL_OP2_RSHIFT, + HLSL_OP3_DP2ADD, HLSL_OP3_LERP, }; diff --git a/libs/vkd3d-shader/hlsl_codegen.c b/libs/vkd3d-shader/hlsl_codegen.c index 4fa860a6..aa950e35 100644 --- a/libs/vkd3d-shader/hlsl_codegen.c +++ b/libs/vkd3d-shader/hlsl_codegen.c @@ -1535,6 +1535,62 @@ static bool lower_sqrt(struct hlsl_ctx *ctx, struct hlsl_ir_node *instr, void *c return true; } +/* Lower DP2 to MUL + ADD */ +static bool lower_dot(struct hlsl_ctx *ctx, struct hlsl_ir_node *instr, void *context) +{ + struct hlsl_ir_node *arg1, *arg2, *mul, *replacement; + struct hlsl_ir_swizzle *add_x, *add_y; + struct hlsl_ir_constant *zero; + struct hlsl_ir_expr *expr; + + if (instr->type != HLSL_IR_EXPR) + return false; + expr = hlsl_ir_expr(instr); + arg1 = expr->operands[0].node; + arg2 = expr->operands[1].node; + if (expr->op != HLSL_OP2_DOT) + return false; + if (arg1->data_type->dimx != 2) + return false; + + if (ctx->profile->type == VKD3D_SHADER_TYPE_PIXEL) + { + struct hlsl_ir_node *operands[HLSL_MAX_OPERANDS] = { 0 }; + + if (!(zero = hlsl_new_float_constant(ctx, 0.0f, &expr->node.loc))) + return false; + list_add_before(&instr->entry, &zero->node.entry); + + operands[0] = arg1; + operands[1] = arg2; + operands[2] = &zero->node; + + if (!(replacement = hlsl_new_expr(ctx, HLSL_OP3_DP2ADD, operands, instr->data_type, &expr->node.loc))) + return false; + } + else + { + if (!(mul = hlsl_new_binary_expr(ctx, HLSL_OP2_MUL, expr->operands[0].node, expr->operands[1].node))) + return false; + list_add_before(&instr->entry, &mul->entry); + + if (!(add_x = hlsl_new_swizzle(ctx, HLSL_SWIZZLE(X, X, X, X), instr->data_type->dimx, mul, &expr->node.loc))) + return false; + list_add_before(&instr->entry, &add_x->node.entry); + + if (!(add_y = hlsl_new_swizzle(ctx, HLSL_SWIZZLE(Y, Y, Y, Y), instr->data_type->dimx, mul, &expr->node.loc))) + return false; + list_add_before(&instr->entry, &add_y->node.entry); + + if (!(replacement = hlsl_new_binary_expr(ctx, HLSL_OP2_ADD, &add_x->node, &add_y->node))) + return false; + } + list_add_before(&instr->entry, &replacement->entry); + + hlsl_replace_node(instr, replacement); + return true; +} + static bool lower_casts_to_bool(struct hlsl_ctx *ctx, struct hlsl_ir_node *instr, void *context) { struct hlsl_type *type = instr->data_type, *arg_type; @@ -2949,6 +3005,7 @@ int hlsl_emit_bytecode(struct hlsl_ctx *ctx, struct hlsl_ir_function_decl *entry { transform_ir(ctx, lower_division, body, NULL); transform_ir(ctx, lower_sqrt, body, NULL); + transform_ir(ctx, lower_dot, body, NULL); } transform_ir(ctx, validate_static_object_references, body, NULL); diff --git a/libs/vkd3d-shader/hlsl_sm1.c b/libs/vkd3d-shader/hlsl_sm1.c index facf81b8..b9fc099d 100644 --- a/libs/vkd3d-shader/hlsl_sm1.c +++ b/libs/vkd3d-shader/hlsl_sm1.c @@ -422,7 +422,7 @@ struct sm1_instruction D3DSHADER_PARAM_SRCMOD_TYPE mod; unsigned int swizzle; uint32_t reg; - } srcs[2]; + } srcs[HLSL_MAX_OPERANDS]; unsigned int src_count; unsigned int has_dst; @@ -459,6 +459,33 @@ static void write_sm1_instruction(struct hlsl_ctx *ctx, struct vkd3d_bytecode_bu write_sm1_src_register(buffer, &instr->srcs[i], instr->dst.writemask); }; +static void write_sm1_ternary_op(struct hlsl_ctx *ctx, struct vkd3d_bytecode_buffer *buffer, + D3DSHADER_INSTRUCTION_OPCODE_TYPE opcode, const struct hlsl_reg *dst, + const struct hlsl_reg *src1, const struct hlsl_reg *src2, const struct hlsl_reg *src3) +{ + const struct sm1_instruction instr = + { + .opcode = opcode, + + .dst.type = D3DSPR_TEMP, + .dst.writemask = dst->writemask, + .dst.reg = dst->id, + .has_dst = 1, + + .srcs[0].type = D3DSPR_TEMP, + .srcs[0].swizzle = hlsl_swizzle_from_writemask(src1->writemask), + .srcs[0].reg = src1->id, + .srcs[1].type = D3DSPR_TEMP, + .srcs[1].swizzle = hlsl_swizzle_from_writemask(src2->writemask), + .srcs[1].reg = src2->id, + .srcs[2].type = D3DSPR_TEMP, + .srcs[2].swizzle = hlsl_swizzle_from_writemask(src3->writemask), + .srcs[2].reg = src3->id, + .src_count = 3, + }; + write_sm1_instruction(ctx, buffer, &instr); +} + static void write_sm1_binary_op(struct hlsl_ctx *ctx, struct vkd3d_bytecode_buffer *buffer, D3DSHADER_INSTRUCTION_OPCODE_TYPE opcode, const struct hlsl_reg *dst, const struct hlsl_reg *src1, const struct hlsl_reg *src2) @@ -631,6 +658,9 @@ static void write_sm1_expr(struct hlsl_ctx *ctx, struct vkd3d_bytecode_buffer *b struct hlsl_ir_expr *expr = hlsl_ir_expr(instr); struct hlsl_ir_node *arg1 = expr->operands[0].node; struct hlsl_ir_node *arg2 = expr->operands[1].node; + struct hlsl_ir_node *arg3 = expr->operands[2].node; + const struct hlsl_type *dst_type = expr->node.data_type; + struct vkd3d_string_buffer *dst_type_string; assert(instr->reg.allocated); @@ -641,6 +671,9 @@ static void write_sm1_expr(struct hlsl_ctx *ctx, struct vkd3d_bytecode_buffer *b return; } + if (!(dst_type_string = hlsl_type_to_string(ctx, dst_type))) + return; + switch (expr->op) { case HLSL_OP1_EXP2: @@ -679,10 +712,32 @@ static void write_sm1_expr(struct hlsl_ctx *ctx, struct vkd3d_bytecode_buffer *b write_sm1_unary_op(ctx, buffer, D3DSIO_FRC, &instr->reg, &arg1->reg, D3DSPSM_NONE); break; + case HLSL_OP2_DOT: + switch (arg1->data_type->dimx) + { + case 4: + write_sm1_binary_op(ctx, buffer, D3DSIO_DP4, &instr->reg, &arg1->reg, &arg2->reg); + break; + + case 3: + write_sm1_binary_op(ctx, buffer, D3DSIO_DP3, &instr->reg, &arg1->reg, &arg2->reg); + break; + + default: + vkd3d_unreachable(); + } + break; + + case HLSL_OP3_DP2ADD: + write_sm1_ternary_op(ctx, buffer, D3DSIO_DP2ADD, &instr->reg, &arg1->reg, &arg2->reg, &arg3->reg); + break; + default: hlsl_fixme(ctx, &instr->loc, "SM1 \"%s\" expression.", debug_hlsl_expr_op(expr->op)); break; } + + hlsl_release_string_buffer(ctx, dst_type_string); } static void write_sm1_load(struct hlsl_ctx *ctx, struct vkd3d_bytecode_buffer *buffer, const struct hlsl_ir_node *instr) -- GitLab https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/74
On Wed Feb 1 17:50:36 2023 +0000, Nikolay Sivov wrote:
changed this line in [version 5 of the diff](/wine/vkd3d/-/merge_requests/74/diffs?diff_id=30394&start_sha=7053503e88ac242f41a14883444ea7a7cab990dc#e23171cc4cf1a77702423c2fd4a3eed6e32c7443_718_718) Thanks, I removed that.
-- https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/74#note_22642
Francisco Casas (@fcasas) commented about libs/vkd3d-shader/hlsl_codegen.c:
+ if (!(replacement = hlsl_new_expr(ctx, HLSL_OP3_DP2ADD, operands, instr->data_type, &expr->node.loc))) + return false; + } + else + { + if (!(mul = hlsl_new_binary_expr(ctx, HLSL_OP2_MUL, expr->operands[0].node, expr->operands[1].node))) + return false; + list_add_before(&instr->entry, &mul->entry); + + if (!(add_x = hlsl_new_swizzle(ctx, HLSL_SWIZZLE(X, X, X, X), instr->data_type->dimx, mul, &expr->node.loc))) + return false; + list_add_before(&instr->entry, &add_x->node.entry); + + if (!(add_y = hlsl_new_swizzle(ctx, HLSL_SWIZZLE(Y, Y, Y, Y), instr->data_type->dimx, mul, &expr->node.loc))) + return false; + list_add_before(&instr->entry, &add_y->node.entry); nitpick: for both calls to `hlsl_new_swizzle()`, `instr->data_type->dimx` could just be `1`, right?
-- https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/74#note_22645
The following shouldn't affect this MR but I found it worth noting: When compiling ```hlsl uniform float2 a, b; float4 main() : sv_target { return dot(a, b); } ``` I see a little discrepancy between how the swizzles are written in native: ``` dp2add r0.x, r0, r1, c2.x ```` versus ours: ``` dp2add r0, r1, r0, r2 ``` The [documentation of `dp2add`](https://learn.microsoft.com/en-us/windows/win32/direct3dhlsl/dp2add---ps) says that the `.x` is a " required replicate swizzle". I have not idea of what that means... Regardless, that swizzle would appear if we just made sure that `arg3->reg.writemask = 1`, but it seems that we are not setting the writemasks to 1 when allocating scalar types in SM1. This may be a bug. -- https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/74#note_22646
Zebediah Figura (@zfigura) commented about libs/vkd3d-shader/hlsl_sm1.c:
D3DSHADER_PARAM_SRCMOD_TYPE mod; unsigned int swizzle; uint32_t reg; - } srcs[2]; + } srcs[HLSL_MAX_OPERANDS];
I wouldn't use HLSL_MAX_OPERANDS here; the maximum number of HLSL expression operands isn't semantically equal to the maximum number of sm1 instruction sources. E.g. sm4 uses 4 here, for things like resource load instructions (I think the actual maximum source count for known instructions is 6). -- https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/74#note_22651
Zebediah Figura (@zfigura) commented about libs/vkd3d-shader/hlsl_sm1.c:
return; }
+ if (!(dst_type_string = hlsl_type_to_string(ctx, dst_type))) + return; +
This isn't being used for anything, is it? -- https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/74#note_22652
participants (4)
-
Francisco Casas (＠fcasas) -
Nikolay Sivov
-
Nikolay Sivov (＠nsivov)
-
Zebediah Figura (＠zfigura)