radeonsi: use compute shader for clear 12-byte buffer

Signed-off-by: Sonny Jiang <[email protected]>
Reviewed-by: Marek Olšák <[email protected]>

4 files changed, 108 insertions, 10 deletions

diff --git a/src/gallium/drivers/radeonsi/si_compute_blit.c b/src/gallium/drivers/radeonsi/si_compute_blit.c
index ff573c131f4..8d4f3bab5d7 100644
--- a/src/gallium/drivers/radeonsi/si_compute_blit.c
+++ b/src/gallium/drivers/radeonsi/si_compute_blit.c
@@ -73,6 +73,76 @@ static void si_compute_internal_end(struct si_context *sctx)
 	sctx->render_cond_force_off = false;
 }
 
+static void si_compute_clear_12bytes_buffer(struct si_context *sctx,
+					struct pipe_resource *dst,
+					unsigned dst_offset,
+					unsigned size,
+					const uint32_t *clear_value,
+					enum si_coherency coher)
+{
+	struct pipe_context *ctx = &sctx->b;
+
+	assert(dst_offset % 4 == 0);
+	assert(size % 4 == 0);
+	unsigned size_12 = DIV_ROUND_UP(size, 12);
+
+	unsigned data[4] = {0};
+	memcpy(data, clear_value, 12);
+
+	si_compute_internal_begin(sctx);
+
+	sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
+		       SI_CONTEXT_CS_PARTIAL_FLUSH |
+		       si_get_flush_flags(sctx, coher, SI_COMPUTE_DST_CACHE_POLICY);
+
+	struct pipe_shader_buffer saved_sb = {0};
+	si_get_shader_buffers(sctx, PIPE_SHADER_COMPUTE, 0, 1, &saved_sb);
+
+	unsigned saved_writable_mask = 0;
+	if (sctx->const_and_shader_buffers[PIPE_SHADER_COMPUTE].writable_mask &
+	    (1u << si_get_shaderbuf_slot(0)))
+		saved_writable_mask = 1;
+
+	struct pipe_constant_buffer saved_cb = {};
+	si_get_pipe_constant_buffer(sctx, PIPE_SHADER_COMPUTE, 0, &saved_cb);
+
+	void *saved_cs = sctx->cs_shader_state.program;
+
+	struct pipe_constant_buffer cb = {};
+	cb.buffer_size = sizeof(data);
+	cb.user_buffer = data;
+	ctx->set_constant_buffer(ctx, PIPE_SHADER_COMPUTE, 0, &cb);
+
+	struct pipe_shader_buffer sb = {0};
+	sb.buffer = dst;
+	sb.buffer_offset = dst_offset;
+	sb.buffer_size = size;
+
+	ctx->set_shader_buffers(ctx, PIPE_SHADER_COMPUTE, 0, 1, &sb, 0x1);
+
+	struct pipe_grid_info info = {0};
+
+	if (!sctx->cs_clear_12bytes_buffer)
+		sctx->cs_clear_12bytes_buffer =
+			si_clear_12bytes_buffer_shader(ctx);
+	ctx->bind_compute_state(ctx, sctx->cs_clear_12bytes_buffer);
+	info.block[0] = 64;
+	info.last_block[0] = size_12 % 64;
+	info.block[1] = 1;
+	info.block[2] = 1;
+	info.grid[0] = DIV_ROUND_UP(size_12, 64);
+	info.grid[1] = 1;
+	info.grid[2] = 1;
+
+	ctx->launch_grid(ctx, &info);
+
+	ctx->bind_compute_state(ctx, saved_cs);
+	ctx->set_shader_buffers(ctx, PIPE_SHADER_COMPUTE, 0, 1, &saved_sb, saved_writable_mask);
+	ctx->set_constant_buffer(ctx, PIPE_SHADER_COMPUTE, 0, &saved_cb);
+
+	si_compute_internal_end(sctx);
+}
+
 static void si_compute_do_clear_or_copy(struct si_context *sctx,
 					struct pipe_resource *dst,
 					unsigned dst_offset,
@@ -231,17 +301,8 @@ void si_clear_buffer(struct si_context *sctx, struct pipe_resource *dst,
 		clear_value_size = 4;
 	}
 
-	/* Use transform feedback for 12-byte clears. */
-	/* TODO: Use compute. */
 	if (clear_value_size == 12) {
-		union pipe_color_union streamout_clear_value;
-
-		memcpy(&streamout_clear_value, clear_value, clear_value_size);
-		si_blitter_begin(sctx, SI_DISABLE_RENDER_COND);
-		util_blitter_clear_buffer(sctx->blitter, dst, offset,
-					  size, clear_value_size / 4,
-					  &streamout_clear_value);
-		si_blitter_end(sctx);
+		si_compute_clear_12bytes_buffer(sctx, dst, offset, size, clear_value, coher);
 		return;
 	}
 
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index aa627279ed3..0ac70a5fdf5 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -232,6 +232,8 @@ static void si_destroy_context(struct pipe_context *context)
 		sctx->b.delete_compute_state(&sctx->b, sctx->cs_clear_render_target);
 	if (sctx->cs_clear_render_target_1d_array)
 		sctx->b.delete_compute_state(&sctx->b, sctx->cs_clear_render_target_1d_array);
+	if (sctx->cs_clear_12bytes_buffer)
+		sctx->b.delete_compute_state(&sctx->b, sctx->cs_clear_12bytes_buffer);
 	if (sctx->cs_dcc_retile)
 		sctx->b.delete_compute_state(&sctx->b, sctx->cs_dcc_retile);
 
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index f313e565d5f..b4e06609582 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -894,6 +894,7 @@ struct si_context {
 	void				*cs_copy_image_1d_array;
 	void				*cs_clear_render_target;
 	void				*cs_clear_render_target_1d_array;
+	void				*cs_clear_12bytes_buffer;
 	void				*cs_dcc_retile;
 	void				*cs_fmask_expand[3][2]; /* [log2(samples)-1][is_array] */
 	struct si_screen		*screen;
@@ -1450,6 +1451,7 @@ void *si_create_copy_image_compute_shader(struct pipe_context *ctx);
 void *si_create_copy_image_compute_shader_1d_array(struct pipe_context *ctx);
 void *si_clear_render_target_shader(struct pipe_context *ctx);
 void *si_clear_render_target_shader_1d_array(struct pipe_context *ctx);
+void *si_clear_12bytes_buffer_shader(struct pipe_context *ctx);
 void *si_create_dcc_retile_cs(struct pipe_context *ctx);
 void *si_create_fmask_expand_cs(struct pipe_context *ctx, unsigned num_samples,
 				bool is_array);
diff --git a/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c b/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c
index 0cf0cd95a8b..90eb39e3506 100644
--- a/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c
+++ b/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c
@@ -665,6 +665,39 @@ void *si_clear_render_target_shader_1d_array(struct pipe_context *ctx)
 	return ctx->create_compute_state(ctx, &state);
 }
 
+void *si_clear_12bytes_buffer_shader(struct pipe_context *ctx)
+{
+	static const char text[] =
+		"COMP\n"
+		"PROPERTY CS_FIXED_BLOCK_WIDTH 64\n"
+		"PROPERTY CS_FIXED_BLOCK_HEIGHT 1\n"
+		"PROPERTY CS_FIXED_BLOCK_DEPTH 1\n"
+		"DCL SV[0], THREAD_ID\n"
+		"DCL SV[1], BLOCK_ID\n"
+		"DCL BUFFER[0]\n"
+		"DCL CONST[0][0..0]\n" // 0:xyzw
+		"DCL TEMP[0..0]\n"
+		"IMM[0] UINT32 {64, 1, 12, 0}\n"
+		"UMAD TEMP[0].x, SV[1].xyzz, IMM[0].xyyy, SV[0].xyzz\n"
+		"UMUL TEMP[0].x, TEMP[0].xyzz, IMM[0].zzzz\n" //12 bytes
+		"STORE BUFFER[0].xyz, TEMP[0].xxxx, CONST[0][0].xyzw\n"
+		"END\n";
+
+	struct tgsi_token tokens[1024];
+	struct pipe_compute_state state = {0};
+
+	if (!tgsi_text_translate(text, tokens, ARRAY_SIZE(tokens))) {
+		assert(false);
+		return NULL;
+	}
+
+	state.ir_type = PIPE_SHADER_IR_TGSI;
+	state.prog = tokens;
+
+	return ctx->create_compute_state(ctx, &state);
+}
+
+
 /* Load samples from the image, and copy them to the same image. This looks like
  * a no-op, but it's not. Loads use FMASK, while stores don't, so samples are
  * reordered to match expanded FMASK.