[Mesa-dev] [PATCH 2/4] radeonsi: switch to using resource destribtors for constants

[Christian König]

From: Christian König <christian.koenig at amd.com>

Signed-off-by: Christian König <christian.koenig at amd.com>
---
 src/gallium/drivers/radeonsi/radeonsi_shader.c |   34 +++++++++++-------------
 src/gallium/drivers/radeonsi/si_state.c        |   32 ++++++++++++++--------
 2 files changed, 37 insertions(+), 29 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/radeonsi_shader.c b/src/gallium/drivers/radeonsi/radeonsi_shader.c
index 95ccd1e..c407953 100644
--- a/src/gallium/drivers/radeonsi/radeonsi_shader.c
+++ b/src/gallium/drivers/radeonsi/radeonsi_shader.c
@@ -354,11 +354,10 @@ static LLVMValueRef fetch_constant(
 {
 	struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
 	struct lp_build_context * base = &bld_base->base;
-	unsigned idx;
 
-	LLVMValueRef const_ptr;
-	LLVMValueRef offset;
-	LLVMValueRef load;
+	LLVMValueRef ptr;
+	LLVMValueRef args[2];
+	LLVMValueRef result;
 
 	if (swizzle == LP_CHAN_ALL) {
 		unsigned chan;
@@ -369,23 +368,22 @@ static LLVMValueRef fetch_constant(
 		return lp_build_gather_values(bld_base->base.gallivm, values, 4);
 	}
 
+	/* Load the resource descriptor */
+	ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_CONST);
+	args[0] = build_indexed_load(base->gallivm, ptr, bld_base->uint_bld.zero);
+
 	/* currently not supported */
 	if (reg->Register.Indirect) {
 		assert(0);
-		load = lp_build_const_int32(base->gallivm, 0);
-		return bitcast(bld_base, type, load);
-	}
-
-	const_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_CONST);
+		result = lp_build_const_int32(base->gallivm, 0);
+		return bitcast(bld_base, type, result);
+	} else
+		args[1] = lp_build_const_int32(base->gallivm, (reg->Register.Index * 4 + swizzle) * 4);
 
-	/* XXX: This assumes that the constant buffer is not packed, so
-	 * CONST[0].x will have an offset of 0 and CONST[1].x will have an
-	 * offset of 4. */
-	idx = (reg->Register.Index * 4) + swizzle;
-	offset = lp_build_const_int32(base->gallivm, idx);
+	result = build_intrinsic(base->gallivm->builder, "llvm.SI.load.const", base->elem_type,
+                                 args, 2, LLVMReadOnlyAttribute | LLVMNoUnwindAttribute);
 
-	load = build_indexed_load(base->gallivm, const_ptr, offset);
-	return bitcast(bld_base, type, load);
+	return bitcast(bld_base, type, result);
 }
 
 /* Initialize arguments for the shader export intrinsic */
@@ -931,8 +929,8 @@ static void create_function(struct si_shader_context *si_shader_ctx)
 	v2i32 = LLVMVectorType(i32, 2);
 	v3i32 = LLVMVectorType(i32, 3);
 
-	params[SI_PARAM_CONST] = LLVMPointerType(f32, CONST_ADDR_SPACE);
-	params[SI_PARAM_SAMPLER] = LLVMPointerType(LLVMVectorType(i8, 16), CONST_ADDR_SPACE);
+	params[SI_PARAM_CONST] = LLVMPointerType(LLVMVectorType(i8, 16), CONST_ADDR_SPACE);
+	params[SI_PARAM_SAMPLER] = params[SI_PARAM_CONST];
 	params[SI_PARAM_RESOURCE] = LLVMPointerType(LLVMVectorType(i8, 32), CONST_ADDR_SPACE);
 
 	if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX) {
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index a395ec4..219add5 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -2480,13 +2480,13 @@ static void si_delete_sampler_state(struct pipe_context *ctx, void *state)
  * Constants
  */
 static void si_set_constant_buffer(struct pipe_context *ctx, uint shader, uint index,
-			    struct pipe_constant_buffer *cb)
+				   struct pipe_constant_buffer *cb)
 {
 	struct r600_context *rctx = (struct r600_context *)ctx;
 	struct si_resource *rbuffer = cb ? si_resource(cb->buffer) : NULL;
 	struct si_pm4_state *pm4;
-	uint64_t va_offset;
-	uint32_t reg, offset;
+	uint32_t offset;
+	uint64_t va;
 
 	/* Note that the state tracker can unbind constant buffers by
 	 * passing NULL here.
@@ -2501,23 +2501,33 @@ static void si_set_constant_buffer(struct pipe_context *ctx, uint shader, uint i
 		r600_upload_const_buffer(rctx, &rbuffer, cb->user_buffer, cb->buffer_size, &offset);
 	else
 		offset = 0;
-	va_offset = r600_resource_va(ctx->screen, (void*)rbuffer);
-	va_offset += offset;
+	va = r600_resource_va(ctx->screen, (void*)rbuffer);
+	va += offset;
 
 	si_pm4_add_bo(pm4, rbuffer, RADEON_USAGE_READ);
 
+	si_pm4_sh_data_begin(pm4);
+
+	/* Fill in a T# buffer resource description */
+	si_pm4_sh_data_add(pm4, va & 0xFFFFFFFF);
+	si_pm4_sh_data_add(pm4, (S_008F04_BASE_ADDRESS_HI(va >> 32) |
+				 S_008F04_STRIDE(0)));
+	si_pm4_sh_data_add(pm4, 0xffffffff); //cb->buffer_size);
+	si_pm4_sh_data_add(pm4, S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
+				S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
+				S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
+				S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
+				S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
+				S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32));
+
 	switch (shader) {
 	case PIPE_SHADER_VERTEX:
-		reg = R_00B130_SPI_SHADER_USER_DATA_VS_0 + SI_SGPR_CONST * 4;
-		si_pm4_set_reg(pm4, reg, va_offset);
-		si_pm4_set_reg(pm4, reg + 4, va_offset >> 32);
+        	si_pm4_sh_data_end(pm4, R_00B130_SPI_SHADER_USER_DATA_VS_0, SI_SGPR_CONST);
 		si_pm4_set_state(rctx, vs_const, pm4);
 		break;
 
 	case PIPE_SHADER_FRAGMENT:
-		reg = R_00B030_SPI_SHADER_USER_DATA_PS_0 + SI_SGPR_CONST * 4;
-		si_pm4_set_reg(pm4, reg, va_offset);
-		si_pm4_set_reg(pm4, reg + 4, va_offset >> 32);
+        	si_pm4_sh_data_end(pm4, R_00B030_SPI_SHADER_USER_DATA_PS_0, SI_SGPR_CONST);
 		si_pm4_set_state(rctx, ps_const, pm4);
 		break;
 
-- 
1.7.9.5