freedreno/a5xx: image support

src/gallium/drivers/freedreno/Makefile.sources

@@ -138,6 +138,8 @@ a5xx_SOURCES := \
 	a5xx/fd5_format.h \
 	a5xx/fd5_gmem.c \
 	a5xx/fd5_gmem.h \
+	a5xx/fd5_image.c \
+	a5xx/fd5_image.h \
 	a5xx/fd5_program.c \
 	a5xx/fd5_program.h \
 	a5xx/fd5_query.c \

src/gallium/drivers/freedreno/a5xx/fd5_emit.c

@@ -37,6 +37,7 @@
 #include "fd5_emit.h"
 #include "fd5_blend.h"
 #include "fd5_context.h"
+#include "fd5_image.h"
 #include "fd5_program.h"
 #include "fd5_rasterizer.h"
 #include "fd5_texture.h"
@@ -764,10 +765,12 @@ fd5_emit_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
 	if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_TEX) {
 		needs_border |= emit_textures(ctx, ring, SB4_FS_TEX,
 				&ctx->tex[PIPE_SHADER_FRAGMENT]);
-		OUT_PKT4(ring, REG_A5XX_TPL1_FS_TEX_COUNT, 1);
-		OUT_RING(ring, ctx->tex[PIPE_SHADER_FRAGMENT].num_textures);
 	}
 
+	OUT_PKT4(ring, REG_A5XX_TPL1_FS_TEX_COUNT, 1);
+	OUT_RING(ring, ctx->shaderimg[PIPE_SHADER_FRAGMENT].enabled_mask ?
+			~0 : ctx->tex[PIPE_SHADER_FRAGMENT].num_textures);
+
 	OUT_PKT4(ring, REG_A5XX_TPL1_CS_TEX_COUNT, 1);
 	OUT_RING(ring, 0);
 
@@ -776,6 +779,9 @@ fd5_emit_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
 
 	if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_SSBO)
 		emit_ssbos(ctx, ring, SB4_SSBO, &ctx->shaderbuf[PIPE_SHADER_FRAGMENT]);
+
+	if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_IMAGE)
+		fd5_emit_images(ctx, ring, PIPE_SHADER_FRAGMENT);
 }
 
 void
@@ -806,13 +812,17 @@ fd5_emit_cs_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
 
 		OUT_PKT4(ring, REG_A5XX_TPL1_FS_TEX_COUNT, 1);
 		OUT_RING(ring, 0);
-
-		OUT_PKT4(ring, REG_A5XX_TPL1_CS_TEX_COUNT, 1);
-		OUT_RING(ring, ctx->tex[PIPE_SHADER_COMPUTE].num_textures);
 	}
 
+	OUT_PKT4(ring, REG_A5XX_TPL1_CS_TEX_COUNT, 1);
+	OUT_RING(ring, ctx->shaderimg[PIPE_SHADER_COMPUTE].enabled_mask ?
+			~0 : ctx->tex[PIPE_SHADER_COMPUTE].num_textures);
+
 	if (dirty & FD_DIRTY_SHADER_SSBO)
 		emit_ssbos(ctx, ring, SB4_CS_SSBO, &ctx->shaderbuf[PIPE_SHADER_COMPUTE]);
+
+	if (dirty & FD_DIRTY_SHADER_IMAGE)
+		fd5_emit_images(ctx, ring, PIPE_SHADER_COMPUTE);
 }
 
 /* emit setup at begin of new cmdstream buffer (don't rely on previous

src/gallium/drivers/freedreno/a5xx/fd5_image.c

@@ -0,0 +1,223 @@
+/*
+ * Copyright (C) 2017 Rob Clark <robclark@freedesktop.org>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ * Authors:
+ *    Rob Clark <robclark@freedesktop.org>
+ */
+
+#include "pipe/p_state.h"
+
+#include "freedreno_resource.h"
+#include "fd5_image.h"
+#include "fd5_format.h"
+#include "fd5_texture.h"
+
+static enum a4xx_state_block texsb[] = {
+	[PIPE_SHADER_COMPUTE] = SB4_CS_TEX,
+	[PIPE_SHADER_FRAGMENT] = SB4_FS_TEX,
+};
+
+static enum a4xx_state_block imgsb[] = {
+	[PIPE_SHADER_COMPUTE] = SB4_CS_SSBO,
+	[PIPE_SHADER_FRAGMENT] = SB4_SSBO,
+};
+
+struct fd5_image {
+	enum pipe_format pfmt;
+	enum a5xx_tex_fmt fmt;
+	enum a5xx_tex_fetchsize fetchsize;
+	enum a5xx_tex_type type;
+	bool srgb;
+	uint32_t cpp;
+	uint32_t width;
+	uint32_t height;
+	uint32_t depth;
+	uint32_t pitch;
+	uint32_t array_pitch;
+	struct fd_bo *bo;
+	uint32_t offset;
+};
+
+static void translate_image(struct fd5_image *img, struct pipe_image_view *pimg)
+{
+	enum pipe_format format = pimg->format;
+	struct pipe_resource *prsc = pimg->resource;
+	struct fd_resource *rsc = fd_resource(prsc);
+	unsigned lvl;
+
+	if (!pimg->resource) {
+		memset(img, 0, sizeof(*img));
+		return;
+	}
+
+	img->pfmt      = format;
+	img->fmt       = fd5_pipe2tex(format);
+	img->fetchsize = fd5_pipe2fetchsize(format);
+	img->type      = fd5_tex_type(prsc->target);
+	img->srgb      = util_format_is_srgb(format);
+	img->cpp       = rsc->cpp;
+	img->bo        = rsc->bo;
+
+	if (prsc->target == PIPE_BUFFER) {
+		lvl = 0;
+		img->offset = pimg->u.buf.offset;
+		img->pitch  = pimg->u.buf.size;
+		img->array_pitch = 0;
+	} else {
+		lvl = pimg->u.tex.level;
+		img->offset = rsc->slices[lvl].offset;
+		img->pitch  = rsc->slices[lvl].pitch * rsc->cpp;
+		img->array_pitch = rsc->layer_size;
+	}
+
+	img->width     = u_minify(prsc->width0, lvl);
+	img->height    = u_minify(prsc->height0, lvl);
+	img->depth     = u_minify(prsc->depth0, lvl);
+}
+
+static void emit_image_tex(struct fd_ringbuffer *ring, unsigned slot,
+		struct fd5_image *img, enum pipe_shader_type shader)
+{
+	OUT_PKT7(ring, CP_LOAD_STATE4, 3 + 12);
+	OUT_RING(ring, CP_LOAD_STATE4_0_DST_OFF(slot) |
+		CP_LOAD_STATE4_0_STATE_SRC(SS4_DIRECT) |
+		CP_LOAD_STATE4_0_STATE_BLOCK(texsb[shader]) |
+		CP_LOAD_STATE4_0_NUM_UNIT(1));
+	OUT_RING(ring, CP_LOAD_STATE4_1_STATE_TYPE(ST4_CONSTANTS) |
+		CP_LOAD_STATE4_1_EXT_SRC_ADDR(0));
+	OUT_RING(ring, CP_LOAD_STATE4_2_EXT_SRC_ADDR_HI(0));
+
+	OUT_RING(ring, A5XX_TEX_CONST_0_FMT(img->fmt) |
+		fd5_tex_swiz(img->pfmt, PIPE_SWIZZLE_X, PIPE_SWIZZLE_Y,
+			PIPE_SWIZZLE_Z, PIPE_SWIZZLE_W) |
+		COND(img->srgb, A5XX_TEX_CONST_0_SRGB));
+	OUT_RING(ring, A5XX_TEX_CONST_1_WIDTH(img->width) |
+		A5XX_TEX_CONST_1_HEIGHT(img->height));
+	OUT_RING(ring, A5XX_TEX_CONST_2_FETCHSIZE(img->fetchsize) |
+		A5XX_TEX_CONST_2_TYPE(img->type) |
+		A5XX_TEX_CONST_2_PITCH(img->pitch));
+	OUT_RING(ring, A5XX_TEX_CONST_3_ARRAY_PITCH(img->array_pitch));
+	if (img->bo) {
+		OUT_RELOC(ring, img->bo, img->offset,
+				(uint64_t)A5XX_TEX_CONST_5_DEPTH(img->depth) << 32, 0);
+	} else {
+		OUT_RING(ring, 0x00000000);
+		OUT_RING(ring, A5XX_TEX_CONST_5_DEPTH(img->depth));
+	}
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+}
+
+static void emit_image_ssbo(struct fd_ringbuffer *ring, unsigned slot,
+		struct fd5_image *img, enum pipe_shader_type shader)
+{
+	OUT_PKT7(ring, CP_LOAD_STATE4, 3 + 4);
+	OUT_RING(ring, CP_LOAD_STATE4_0_DST_OFF(slot) |
+		CP_LOAD_STATE4_0_STATE_SRC(SS4_DIRECT) |
+		CP_LOAD_STATE4_0_STATE_BLOCK(imgsb[shader]) |
+		CP_LOAD_STATE4_0_NUM_UNIT(1));
+	OUT_RING(ring, CP_LOAD_STATE4_1_STATE_TYPE(0) |
+		CP_LOAD_STATE4_1_EXT_SRC_ADDR(0));
+	OUT_RING(ring, CP_LOAD_STATE4_2_EXT_SRC_ADDR_HI(0));
+	OUT_RING(ring, A5XX_SSBO_0_0_BASE_LO(0));
+	OUT_RING(ring, A5XX_SSBO_0_1_PITCH(img->pitch));
+	OUT_RING(ring, A5XX_SSBO_0_2_ARRAY_PITCH(img->array_pitch));
+	OUT_RING(ring, A5XX_SSBO_0_3_CPP(img->cpp));
+
+	OUT_PKT7(ring, CP_LOAD_STATE4, 3 + 2);
+	OUT_RING(ring, CP_LOAD_STATE4_0_DST_OFF(slot) |
+		CP_LOAD_STATE4_0_STATE_SRC(SS4_DIRECT) |
+		CP_LOAD_STATE4_0_STATE_BLOCK(imgsb[shader]) |
+		CP_LOAD_STATE4_0_NUM_UNIT(1));
+	OUT_RING(ring, CP_LOAD_STATE4_1_STATE_TYPE(1) |
+		CP_LOAD_STATE4_1_EXT_SRC_ADDR(0));
+	OUT_RING(ring, CP_LOAD_STATE4_2_EXT_SRC_ADDR_HI(0));
+	OUT_RING(ring, A5XX_SSBO_1_0_FMT(img->fmt) |
+		A5XX_SSBO_1_0_WIDTH(img->width));
+	OUT_RING(ring, A5XX_SSBO_1_1_HEIGHT(img->height) |
+		A5XX_SSBO_1_1_DEPTH(img->depth));
+
+	OUT_PKT7(ring, CP_LOAD_STATE4, 3 + 2);
+	OUT_RING(ring, CP_LOAD_STATE4_0_DST_OFF(slot) |
+		CP_LOAD_STATE4_0_STATE_SRC(SS4_DIRECT) |
+		CP_LOAD_STATE4_0_STATE_BLOCK(imgsb[shader]) |
+		CP_LOAD_STATE4_0_NUM_UNIT(1));
+	OUT_RING(ring, CP_LOAD_STATE4_1_STATE_TYPE(2) |
+		CP_LOAD_STATE4_1_EXT_SRC_ADDR(0));
+	OUT_RING(ring, CP_LOAD_STATE4_2_EXT_SRC_ADDR_HI(0));
+	if (img->bo) {
+		OUT_RELOCW(ring, img->bo, img->offset, 0, 0);
+	} else {
+		OUT_RING(ring, 0x00000000);
+		OUT_RING(ring, 0x00000000);
+	}
+}
+
+/* Note that to avoid conflicts with textures and non-image "SSBO"s, images
+ * are placedd, in reverse order, at the end of the state block, so for
+ * example the sampler state:
+ *
+ *   0:   first texture
+ *   1:   second texture
+ *   ....
+ *   N-1: second image
+ *   N:   first image
+ */
+static unsigned
+get_image_slot(unsigned index)
+{
+	/* TODO figure out real limit per generation, and don't hardcode.
+	 * This needs to match get_image_slot() in ir3_compiler_nir.
+	 * Possibly should be factored out into shared helper?
+	 */
+	const unsigned max_samplers = 16;
+	return max_samplers - index - 1;
+}
+
+/* Emit required "SSBO" and sampler state.  The sampler state is used by the
+ * hw for imageLoad(), and "SSBO" state for imageStore().  Returns max sampler
+ * used.
+ */
+void
+fd5_emit_images(struct fd_context *ctx, struct fd_ringbuffer *ring,
+		enum pipe_shader_type shader)
+{
+	struct fd_shaderimg_stateobj *so = &ctx->shaderimg[shader];
+
+	so->dirty_mask &= so->enabled_mask;
+
+	while (so->dirty_mask) {
+		unsigned index = u_bit_scan(&so->dirty_mask);
+		unsigned slot = get_image_slot(index);
+		struct fd5_image img;
+
+		translate_image(&img, &so->si[index]);
+
+		emit_image_tex(ring, slot, &img, shader);
+		emit_image_ssbo(ring, slot, &img, shader);
+	}
+}
+

src/gallium/drivers/freedreno/a5xx/fd5_image.h

@@ -0,0 +1,35 @@
+/*
+ * Copyright (C) 2017 Rob Clark <robclark@freedesktop.org>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ * Authors:
+ *    Rob Clark <robclark@freedesktop.org>
+ */
+
+#ifndef FD5_IMAGE_H_
+#define FD5_IMAGE_H_
+
+#include "freedreno_context.h"
+
+void fd5_emit_images(struct fd_context *ctx, struct fd_ringbuffer *ring,
+		enum pipe_shader_type shader);
+
+#endif /* FD5_IMAGE_H_ */

src/gallium/drivers/freedreno/a5xx/fd5_texture.c

@@ -186,28 +186,6 @@ fd5_sampler_states_bind(struct pipe_context *pctx,
 	}
 }
 
-static enum a5xx_tex_type
-tex_type(unsigned target)
-{
-	switch (target) {
-	default:
-		assert(0);
-	case PIPE_BUFFER:
-	case PIPE_TEXTURE_1D:
-	case PIPE_TEXTURE_1D_ARRAY:
-		return A5XX_TEX_1D;
-	case PIPE_TEXTURE_RECT:
-	case PIPE_TEXTURE_2D:
-	case PIPE_TEXTURE_2D_ARRAY:
-		return A5XX_TEX_2D;
-	case PIPE_TEXTURE_3D:
-		return A5XX_TEX_3D;
-	case PIPE_TEXTURE_CUBE:
-	case PIPE_TEXTURE_CUBE_ARRAY:
-		return A5XX_TEX_CUBE;
-	}
-}
-
 static bool
 use_astc_srgb_workaround(struct pipe_context *pctx, enum pipe_format format)
 {
@@ -272,7 +250,7 @@ fd5_sampler_view_create(struct pipe_context *pctx, struct pipe_resource *prsc,
 		so->offset = fd_resource_offset(rsc, lvl, cso->u.tex.first_layer);
 	}
 
-	so->texconst2 |= A5XX_TEX_CONST_2_TYPE(tex_type(cso->target));
+	so->texconst2 |= A5XX_TEX_CONST_2_TYPE(fd5_tex_type(cso->target));
 
 	switch (cso->target) {
 	case PIPE_TEXTURE_1D:

src/gallium/drivers/freedreno/a5xx/fd5_texture.h

@@ -67,4 +67,27 @@ unsigned fd5_get_const_idx(struct fd_context *ctx,
 
 void fd5_texture_init(struct pipe_context *pctx);
 
+
+static inline enum a5xx_tex_type
+fd5_tex_type(unsigned target)
+{
+	switch (target) {
+	default:
+		assert(0);
+	case PIPE_BUFFER:
+	case PIPE_TEXTURE_1D:
+	case PIPE_TEXTURE_1D_ARRAY:
+		return A5XX_TEX_1D;
+	case PIPE_TEXTURE_RECT:
+	case PIPE_TEXTURE_2D:
+	case PIPE_TEXTURE_2D_ARRAY:
+		return A5XX_TEX_2D;
+	case PIPE_TEXTURE_3D:
+		return A5XX_TEX_3D;
+	case PIPE_TEXTURE_CUBE:
+	case PIPE_TEXTURE_CUBE_ARRAY:
+		return A5XX_TEX_CUBE;
+	}
+}
+
 #endif /* FD5_TEXTURE_H_ */

src/gallium/drivers/freedreno/freedreno_screen.c

@@ -558,6 +558,7 @@ fd_screen_get_shader_param(struct pipe_screen *pscreen,
 	case PIPE_SHADER_CAP_MAX_HW_ATOMIC_COUNTER_BUFFERS:
 		return 0;
 	case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
+	case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
 		if (is_a5xx(screen)) {
 			/* a5xx (and a4xx for that matter) has one state-block
 			 * for compute-shader SSBO's and another that is shared
@@ -576,6 +577,10 @@ fd_screen_get_shader_param(struct pipe_screen *pscreen,
 			 *
 			 * I think that way we could avoid having to patch shaders
 			 * for actual SSBO indexes by using a static partitioning.
+			 *
+			 * Note same state block is used for images and buffers,
+			 * but images also need texture state for read access
+			 * (isam/isam.3d)
 			 */
 			switch(shader)
 			{
@@ -587,9 +592,6 @@ fd_screen_get_shader_param(struct pipe_screen *pscreen,
 			}
 		}
 		return 0;
-	case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
-		/* probably should be same as MAX_SHADRER_BUFFERS but not implemented yet */
-		return 0;
 	}
 	debug_printf("unknown shader param %d\n", param);
 	return 0;

src/gallium/drivers/freedreno/meson.build

@@ -160,6 +160,8 @@ files_libfreedreno = files(
   'a5xx/fd5_format.h',
   'a5xx/fd5_gmem.c',
   'a5xx/fd5_gmem.h',
+  'a5xx/fd5_image.c',
+  'a5xx/fd5_image.h',
   'a5xx/fd5_program.c',
   'a5xx/fd5_program.h',
   'a5xx/fd5_query.c',