From e054251ed11e25a080f64b92db9334c9b07c8c76 Mon Sep 17 00:00:00 2001 From: Brian Paul Date: Thu, 13 Aug 2015 11:00:58 -0700 Subject: svga: update driver for version 10 GPU interface This is a squash commit of roughly two years of development work. Authors include: Brian Paul Charmaine Lee Thomas Hellstrom Jakob Bornecrantz Sinclair Yeh Mingcheng Chen Kai Ninomiya MengLin Wu The driver supports OpenGL 3.3. Signed-off-by: Brian Paul --- src/gallium/drivers/svga/svga_state_constants.c | 588 +++++++++++++++++++----- 1 file changed, 474 insertions(+), 114 deletions(-) (limited to 'src/gallium/drivers/svga/svga_state_constants.c') diff --git a/src/gallium/drivers/svga/svga_state_constants.c b/src/gallium/drivers/svga/svga_state_constants.c index 1e1fbb099c6..e0bbe013c02 100644 --- a/src/gallium/drivers/svga/svga_state_constants.c +++ b/src/gallium/drivers/svga/svga_state_constants.c @@ -1,3 +1,4 @@ + /********************************************************** * Copyright 2008-2009 VMware, Inc. All rights reserved. * @@ -23,9 +24,11 @@ * **********************************************************/ +#include "util/u_format.h" #include "util/u_inlines.h" #include "util/u_memory.h" #include "pipe/p_defines.h" +#include "util/u_upload_mgr.h" #include "svga_screen.h" #include "svga_context.h" @@ -34,6 +37,7 @@ #include "svga_tgsi.h" #include "svga_debug.h" #include "svga_resource_buffer.h" +#include "svga_shader.h" #include "svga_hw_reg.h" @@ -52,65 +56,135 @@ /** Guest-backed surface constant buffers must be this size */ #define GB_CONSTBUF_SIZE (SVGA3D_CONSTREG_MAX) + /** - * Convert from PIPE_SHADER_* to SVGA3D_SHADERTYPE_* + * Emit any extra shader-type-independent shader constants into the buffer + * pointed to by 'dest'. + * \return number of float[4] constants put into the 'dest' buffer */ static unsigned -svga_shader_type(unsigned shader) +svga_get_extra_constants_common(struct svga_context *svga, + const struct svga_shader_variant *variant, + unsigned shader, float *dest) { - switch (shader) { - case PIPE_SHADER_VERTEX: - return SVGA3D_SHADERTYPE_VS; - case PIPE_SHADER_FRAGMENT: - return SVGA3D_SHADERTYPE_PS; - default: - assert(!"Unexpected shader type"); - return SVGA3D_SHADERTYPE_VS; + uint32_t *dest_u = (uint32_t *) dest; // uint version of dest + unsigned i; + unsigned count = 0; + + for (i = 0; i < variant->key.num_textures; i++) { + struct pipe_sampler_view *sv = svga->curr.sampler_views[shader][i]; + if (sv) { + struct pipe_resource *tex = sv->texture; + /* Scaling factors needed for handling unnormalized texture coordinates + * for texture rectangles. + */ + if (variant->key.tex[i].unnormalized) { + /* debug/sanity check */ + assert(variant->key.tex[i].width_height_idx == count); + + *dest++ = 1.0 / (float)tex->width0; + *dest++ = 1.0 / (float)tex->height0; + *dest++ = 1.0; + *dest++ = 1.0; + + count++; + } + + /* Store the sizes for texture buffers. + */ + if (tex->target == PIPE_BUFFER) { + unsigned bytes_per_element = util_format_get_blocksize(sv->format); + *dest_u++ = tex->width0 / bytes_per_element; + *dest_u++ = 1; + *dest_u++ = 1; + *dest_u++ = 1; + + count++; + } + } } + + return count; } /** * Emit any extra fragment shader constants into the buffer pointed * to by 'dest'. - * In particular, these would be the scaling factors needed for handling - * unnormalized texture coordinates for texture rectangles. * \return number of float[4] constants put into the dest buffer */ static unsigned svga_get_extra_fs_constants(struct svga_context *svga, float *dest) { const struct svga_shader_variant *variant = svga->state.hw_draw.fs; - const struct svga_fs_compile_key *key = &variant->key.fkey; unsigned count = 0; - /* SVGA_NEW_VS_VARIANT - */ - if (key->num_unnormalized_coords) { - unsigned i; + count += svga_get_extra_constants_common(svga, variant, + PIPE_SHADER_FRAGMENT, dest); - for (i = 0; i < key->num_textures; i++) { - if (key->tex[i].unnormalized) { - struct pipe_resource *tex = svga->curr.sampler_views[i]->texture; + assert(count <= MAX_EXTRA_CONSTS); - /* debug/sanity check */ - assert(key->tex[i].width_height_idx == count); + return count; +} - *dest++ = 1.0 / (float)tex->width0; - *dest++ = 1.0 / (float)tex->height0; - *dest++ = 1.0; - *dest++ = 1.0; +/** + * Emit extra constants needed for prescale computation into the + * the buffer pointed to by '*dest'. The updated buffer pointer + * will be returned in 'dest'. + */ +static unsigned +svga_get_prescale_constants(struct svga_context *svga, float **dest) +{ + memcpy(*dest, svga->state.hw_clear.prescale.scale, 4 * sizeof(float)); + *dest += 4; - count++; - } - } - } + memcpy(*dest, svga->state.hw_clear.prescale.translate, 4 * sizeof(float)); + *dest += 4; - assert(count <= MAX_EXTRA_CONSTS); + return 2; +} - return count; +/** + * Emit extra constants needed for point sprite emulation. + */ +static unsigned +svga_get_pt_sprite_constants(struct svga_context *svga, float **dest) +{ + struct svga_screen *screen = svga_screen(svga->pipe.screen); + float *dst = *dest; + + dst[0] = 1.0 / (svga->curr.viewport.scale[0] * 2); + dst[1] = 1.0 / (svga->curr.viewport.scale[1] * 2); + dst[2] = svga->curr.rast->pointsize; + dst[3] = screen->maxPointSize; + *dest = *dest + 4; + return 1; } +/** + * Emit user-defined clip plane coefficients into the buffer pointed to + * by '*dest'. The updated buffer pointer will be returned in 'dest'. + */ +static unsigned +svga_get_clip_plane_constants(struct svga_context *svga, + const struct svga_shader_variant *variant, + float **dest) +{ + unsigned count = 0; + + /* SVGA_NEW_CLIP */ + if (svga_have_vgpu10(svga)) { + /* append user-defined clip plane coefficients onto constant buffer */ + unsigned clip_planes = variant->key.clip_plane_enable; + while (clip_planes) { + int i = u_bit_scan(&clip_planes); + COPY_4V(*dest, svga->curr.clip.ucp[i]); + *dest += 4; + count += 1; + } + } + return count; +} /** * Emit any extra vertex shader constants into the buffer pointed @@ -124,26 +198,71 @@ static unsigned svga_get_extra_vs_constants(struct svga_context *svga, float *dest) { const struct svga_shader_variant *variant = svga->state.hw_draw.vs; - const struct svga_vs_compile_key *key = &variant->key.vkey; unsigned count = 0; /* SVGA_NEW_VS_VARIANT */ - if (key->need_prescale) { - memcpy(dest, svga->state.hw_clear.prescale.scale, 4 * sizeof(float)); - dest += 4; + if (variant->key.vs.need_prescale) { + count += svga_get_prescale_constants(svga, &dest); + } - memcpy(dest, svga->state.hw_clear.prescale.translate, 4 * sizeof(float)); + if (variant->key.vs.undo_viewport) { + /* Used to convert window coords back to NDC coords */ + dest[0] = 1.0f / svga->curr.viewport.scale[0]; + dest[1] = 1.0f / svga->curr.viewport.scale[1]; + dest[2] = -svga->curr.viewport.translate[0]; + dest[3] = -svga->curr.viewport.translate[1]; dest += 4; - - count = 2; + count += 1; } + /* SVGA_NEW_CLIP */ + count += svga_get_clip_plane_constants(svga, variant, &dest); + + /* common constants */ + count += svga_get_extra_constants_common(svga, variant, + PIPE_SHADER_VERTEX, dest); + assert(count <= MAX_EXTRA_CONSTS); return count; } +/** + * Emit any extra geometry shader constants into the buffer pointed + * to by 'dest'. + */ +static unsigned +svga_get_extra_gs_constants(struct svga_context *svga, float *dest) +{ + const struct svga_shader_variant *variant = svga->state.hw_draw.gs; + unsigned count = 0; + + /* SVGA_NEW_GS_VARIANT + */ + + /* Constants for point sprite + * These are used in the transformed gs that supports point sprite. + * They need to be added before the prescale constants. + */ + if (variant->key.gs.wide_point) { + count += svga_get_pt_sprite_constants(svga, &dest); + } + + if (variant->key.gs.need_prescale) { + count += svga_get_prescale_constants(svga, &dest); + } + + /* SVGA_NEW_CLIP */ + count += svga_get_clip_plane_constants(svga, variant, &dest); + + /* common constants */ + count += svga_get_extra_constants_common(svga, variant, + PIPE_SHADER_GEOMETRY, dest); + + assert(count <= MAX_EXTRA_CONSTS); + return count; +} /** * Check and emit one shader constant register. @@ -159,6 +278,7 @@ emit_const(struct svga_context *svga, unsigned shader, unsigned i, assert(shader < PIPE_SHADER_TYPES); assert(i < SVGA3D_CONSTREG_MAX); + assert(!svga_have_vgpu10(svga)); if (memcmp(svga->state.hw_draw.cb[shader][i], value, 4 * sizeof(float)) != 0) { @@ -202,6 +322,10 @@ emit_const_range(struct svga_context *svga, unsigned i, j; enum pipe_error ret; + assert(shader == PIPE_SHADER_VERTEX || + shader == PIPE_SHADER_FRAGMENT); + assert(!svga_have_vgpu10(svga)); + #ifdef DEBUG if (offset + count > SVGA3D_CONSTREG_MAX) { debug_printf("svga: too many constants (offset %u + count %u = %u (max = %u))\n", @@ -307,10 +431,12 @@ emit_const_range(struct svga_context *svga, /** * Emit all the constants in a constant buffer for a shader stage. + * On VGPU10, emit_consts_vgpu10 is used instead. */ static enum pipe_error -emit_consts(struct svga_context *svga, unsigned shader) +emit_consts_vgpu9(struct svga_context *svga, unsigned shader) { + const struct pipe_constant_buffer *cbuf; struct svga_screen *ss = svga_screen(svga->pipe.screen); struct pipe_transfer *transfer = NULL; unsigned count; @@ -320,53 +446,284 @@ emit_consts(struct svga_context *svga, unsigned shader) const unsigned offset = 0; assert(shader < PIPE_SHADER_TYPES); + assert(!svga_have_vgpu10(svga)); + /* Only one constant buffer per shader is supported before VGPU10. + * This is only an approximate check against that. + */ + assert(svga->curr.constbufs[shader][1].buffer == NULL); - if (svga->curr.cbufs[shader].buffer == NULL) - goto done; + cbuf = &svga->curr.constbufs[shader][0]; - data = (const float (*)[4])pipe_buffer_map(&svga->pipe, - svga->curr.cbufs[shader].buffer, - PIPE_TRANSFER_READ, - &transfer); - if (data == NULL) { - ret = PIPE_ERROR_OUT_OF_MEMORY; - goto done; - } + if (svga->curr.constbufs[shader][0].buffer) { + /* emit user-provided constants */ + data = (const float (*)[4]) + pipe_buffer_map(&svga->pipe, svga->curr.constbufs[shader][0].buffer, + PIPE_TRANSFER_READ, &transfer); + if (data == NULL) { + return PIPE_ERROR_OUT_OF_MEMORY; + } - /* sanity check */ - assert(svga->curr.cbufs[shader].buffer->width0 >= - svga->curr.cbufs[shader].buffer_size); + /* sanity check */ + assert(cbuf->buffer->width0 >= + cbuf->buffer_size); - /* Use/apply the constant buffer size and offsets here */ - count = svga->curr.cbufs[shader].buffer_size / (4 * sizeof(float)); - data += svga->curr.cbufs[shader].buffer_offset / (4 * sizeof(float)); + /* Use/apply the constant buffer size and offsets here */ + count = cbuf->buffer_size / (4 * sizeof(float)); + data += cbuf->buffer_offset / (4 * sizeof(float)); + + if (ss->hw_version >= SVGA3D_HWVERSION_WS8_B1) { + ret = emit_const_range( svga, shader, offset, count, data ); + } + else { + for (i = 0; i < count; i++) { + ret = emit_const( svga, shader, offset + i, data[i] ); + if (ret != PIPE_OK) { + break; + } + } + } + + pipe_buffer_unmap(&svga->pipe, transfer); - if (ss->hw_version >= SVGA3D_HWVERSION_WS8_B1) { - ret = emit_const_range( svga, shader, offset, count, data ); if (ret != PIPE_OK) { - goto done; + return ret; } - } else { - for (i = 0; i < count; i++) { - ret = emit_const( svga, shader, offset + i, data[i] ); - if (ret != PIPE_OK) { - goto done; + } + + /* emit extra shader constants */ + { + const struct svga_shader_variant *variant = NULL; + unsigned offset; + float extras[MAX_EXTRA_CONSTS][4]; + unsigned count, i; + + switch (shader) { + case PIPE_SHADER_VERTEX: + variant = svga->state.hw_draw.vs; + count = svga_get_extra_vs_constants(svga, (float *) extras); + break; + case PIPE_SHADER_FRAGMENT: + variant = svga->state.hw_draw.fs; + count = svga_get_extra_fs_constants(svga, (float *) extras); + break; + default: + assert(!"Unexpected shader type"); + count = 0; + } + + assert(variant); + offset = variant->shader->info.file_max[TGSI_FILE_CONSTANT] + 1; + assert(count <= Elements(extras)); + + if (count > 0) { + if (ss->hw_version >= SVGA3D_HWVERSION_WS8_B1) { + ret = emit_const_range(svga, shader, offset, count, + (const float (*) [4])extras); + } + else { + for (i = 0; i < count; i++) { + ret = emit_const(svga, shader, offset + i, extras[i]); + if (ret != PIPE_OK) + return ret; + } } } } -done: - if (data) - pipe_buffer_unmap(&svga->pipe, transfer); + return ret; +} + + + +static enum pipe_error +emit_constbuf_vgpu10(struct svga_context *svga, unsigned shader) +{ + const struct pipe_constant_buffer *cbuf; + struct pipe_resource *dst_buffer = NULL; + enum pipe_error ret = PIPE_OK; + struct pipe_transfer *src_transfer; + struct svga_winsys_surface *dst_handle; + float extras[MAX_EXTRA_CONSTS][4]; + unsigned extra_count, extra_size, extra_offset; + unsigned new_buf_size; + void *src_map = NULL, *dst_map; + unsigned offset; + const struct svga_shader_variant *variant; + + assert(shader == PIPE_SHADER_VERTEX || + shader == PIPE_SHADER_GEOMETRY || + shader == PIPE_SHADER_FRAGMENT); + + cbuf = &svga->curr.constbufs[shader][0]; + + switch (shader) { + case PIPE_SHADER_VERTEX: + variant = svga->state.hw_draw.vs; + extra_count = svga_get_extra_vs_constants(svga, (float *) extras); + break; + case PIPE_SHADER_FRAGMENT: + variant = svga->state.hw_draw.fs; + extra_count = svga_get_extra_fs_constants(svga, (float *) extras); + break; + case PIPE_SHADER_GEOMETRY: + variant = svga->state.hw_draw.gs; + extra_count = svga_get_extra_gs_constants(svga, (float *) extras); + break; + default: + assert(!"Unexpected shader type"); + /* Don't return an error code since we don't want to keep re-trying + * this function and getting stuck in an infinite loop. + */ + return PIPE_OK; + } + + assert(variant); + + /* Compute extra constants size and offset in bytes */ + extra_size = extra_count * 4 * sizeof(float); + extra_offset = 4 * sizeof(float) * variant->extra_const_start; + + if (cbuf->buffer_size + extra_size == 0) + return PIPE_OK; /* nothing to do */ + + /* Typically, the cbuf->buffer here is a user-space buffer so mapping + * it is really cheap. If we ever get real HW buffers for constants + * we should void mapping and instead use a ResourceCopy command. + */ + if (cbuf->buffer_size > 0) { + src_map = pipe_buffer_map_range(&svga->pipe, cbuf->buffer, + cbuf->buffer_offset, cbuf->buffer_size, + PIPE_TRANSFER_READ, &src_transfer); + assert(src_map); + if (!src_map) { + return PIPE_ERROR_OUT_OF_MEMORY; + } + } + + /* The new/dest buffer's size must be large enough to hold the original, + * user-specified constants, plus the extra constants. + * The size of the original constant buffer _should_ agree with what the + * shader is expecting, but it might not (it's not enforced anywhere by + * gallium). + */ + new_buf_size = MAX2(cbuf->buffer_size, extra_offset) + extra_size; + + /* According to the DX10 spec, the constant buffer size must be + * in multiples of 16. + */ + new_buf_size = align(new_buf_size, 16); + + ret = u_upload_alloc(svga->const0_upload, 0, new_buf_size, &offset, + &dst_buffer, &dst_map); + if (ret != PIPE_OK || !dst_map) { + if (src_map) + pipe_buffer_unmap(&svga->pipe, src_transfer); + return PIPE_ERROR_OUT_OF_MEMORY; + } + + if (src_map) { + memcpy(dst_map, src_map, cbuf->buffer_size); + pipe_buffer_unmap(&svga->pipe, src_transfer); + } + + if (extra_size) { + assert(extra_offset + extra_size <= new_buf_size); + memcpy((char *) dst_map + extra_offset, extras, extra_size); + } + u_upload_unmap(svga->const0_upload); + + /* Issue the SetSingleConstantBuffer command */ + dst_handle = svga_buffer_handle(svga, dst_buffer); + if (!dst_handle) { + pipe_resource_reference(&dst_buffer, NULL); + return PIPE_ERROR_OUT_OF_MEMORY; + } + + assert(new_buf_size % 16 == 0); + ret = SVGA3D_vgpu10_SetSingleConstantBuffer(svga->swc, + 0, /* index */ + svga_shader_type(shader), + dst_handle, + offset, + new_buf_size); + + if (ret != PIPE_OK) { + pipe_resource_reference(&dst_buffer, NULL); + return ret; + } + + /* Save this const buffer until it's replaced in the future. + * Otherwise, all references to the buffer will go away after the + * command buffer is submitted, it'll get recycled and we will have + * incorrect constant buffer bindings. + */ + pipe_resource_reference(&svga->state.hw_draw.constbuf[shader], dst_buffer); + + svga->state.hw_draw.default_constbuf_size[shader] = new_buf_size; + + pipe_resource_reference(&dst_buffer, NULL); return ret; } +static enum pipe_error +emit_consts_vgpu10(struct svga_context *svga, unsigned shader) +{ + enum pipe_error ret; + unsigned dirty_constbufs; + unsigned enabled_constbufs; + + /* Emit 0th constant buffer (with extra constants) */ + ret = emit_constbuf_vgpu10(svga, shader); + if (ret != PIPE_OK) { + return ret; + } + + enabled_constbufs = svga->state.hw_draw.enabled_constbufs[shader] | 1u; + + /* Emit other constant buffers (UBOs) */ + dirty_constbufs = svga->state.dirty_constbufs[shader] & ~1u; + + while (dirty_constbufs) { + unsigned index = u_bit_scan(&dirty_constbufs); + unsigned offset = svga->curr.constbufs[shader][index].buffer_offset; + unsigned size = svga->curr.constbufs[shader][index].buffer_size; + struct svga_buffer *buffer = + svga_buffer(svga->curr.constbufs[shader][index].buffer); + struct svga_winsys_surface *handle; + + if (buffer) { + handle = svga_buffer_handle(svga, &buffer->b.b); + enabled_constbufs |= 1 << index; + } + else { + handle = NULL; + enabled_constbufs &= ~(1 << index); + assert(offset == 0); + assert(size == 0); + } + + assert(size % 16 == 0); + ret = SVGA3D_vgpu10_SetSingleConstantBuffer(svga->swc, + index, + svga_shader_type(shader), + handle, + offset, + size); + if (ret != PIPE_OK) + return ret; + } + + svga->state.hw_draw.enabled_constbufs[shader] = enabled_constbufs; + svga->state.dirty_constbufs[shader] = 0; + + return ret; +} + static enum pipe_error emit_fs_consts(struct svga_context *svga, unsigned dirty) { - struct svga_screen *ss = svga_screen(svga->pipe.screen); const struct svga_shader_variant *variant = svga->state.hw_draw.fs; enum pipe_error ret = PIPE_OK; @@ -377,28 +734,11 @@ emit_fs_consts(struct svga_context *svga, unsigned dirty) /* SVGA_NEW_FS_CONST_BUFFER */ - ret = emit_consts( svga, PIPE_SHADER_FRAGMENT ); - if (ret != PIPE_OK) - return ret; - - /* emit extra shader constants */ - { - unsigned offset = variant->shader->info.file_max[TGSI_FILE_CONSTANT] + 1; - float extras[MAX_EXTRA_CONSTS][4]; - unsigned count, i; - - count = svga_get_extra_fs_constants(svga, (float *) extras); - - if (ss->hw_version >= SVGA3D_HWVERSION_WS8_B1) { - ret = emit_const_range(svga, PIPE_SHADER_FRAGMENT, offset, count, - (const float (*) [4])extras); - } else { - for (i = 0; i < count; i++) { - ret = emit_const(svga, PIPE_SHADER_FRAGMENT, offset + i, extras[i]); - if (ret != PIPE_OK) - return ret; - } - } + if (svga_have_vgpu10(svga)) { + ret = emit_consts_vgpu10(svga, PIPE_SHADER_FRAGMENT); + } + else { + ret = emit_consts_vgpu9(svga, PIPE_SHADER_FRAGMENT); } return ret; @@ -419,7 +759,6 @@ struct svga_tracked_state svga_hw_fs_constants = static enum pipe_error emit_vs_consts(struct svga_context *svga, unsigned dirty) { - struct svga_screen *ss = svga_screen(svga->pipe.screen); const struct svga_shader_variant *variant = svga->state.hw_draw.vs; enum pipe_error ret = PIPE_OK; @@ -430,29 +769,11 @@ emit_vs_consts(struct svga_context *svga, unsigned dirty) /* SVGA_NEW_VS_CONST_BUFFER */ - ret = emit_consts( svga, PIPE_SHADER_VERTEX ); - if (ret != PIPE_OK) - return ret; - - /* emit extra shader constants */ - { - unsigned offset = variant->shader->info.file_max[TGSI_FILE_CONSTANT] + 1; - float extras[MAX_EXTRA_CONSTS][4]; - unsigned count, i; - - count = svga_get_extra_vs_constants(svga, (float *) extras); - assert(count <= Elements(extras)); - - if (ss->hw_version >= SVGA3D_HWVERSION_WS8_B1) { - ret = emit_const_range(svga, PIPE_SHADER_VERTEX, offset, count, - (const float (*) [4]) extras); - } else { - for (i = 0; i < count; i++) { - ret = emit_const(svga, PIPE_SHADER_VERTEX, offset + i, extras[i]); - if (ret != PIPE_OK) - return ret; - } - } + if (svga_have_vgpu10(svga)) { + ret = emit_consts_vgpu10(svga, PIPE_SHADER_VERTEX); + } + else { + ret = emit_consts_vgpu9(svga, PIPE_SHADER_VERTEX); } return ret; @@ -467,3 +788,42 @@ struct svga_tracked_state svga_hw_vs_constants = SVGA_NEW_VS_VARIANT), emit_vs_consts }; + + +static enum pipe_error +emit_gs_consts(struct svga_context *svga, unsigned dirty) +{ + const struct svga_shader_variant *variant = svga->state.hw_draw.gs; + enum pipe_error ret = PIPE_OK; + + /* SVGA_NEW_GS_VARIANT + */ + if (variant == NULL) + return PIPE_OK; + + /* SVGA_NEW_GS_CONST_BUFFER + */ + if (svga_have_vgpu10(svga)) { + /** + * If only the rasterizer state has changed and the current geometry + * shader does not emit wide points, then there is no reason to + * re-emit the GS constants, so skip it. + */ + if (dirty == SVGA_NEW_RAST && !variant->key.gs.wide_point) + return PIPE_OK; + + ret = emit_consts_vgpu10(svga, PIPE_SHADER_GEOMETRY); + } + + return ret; +} + + +struct svga_tracked_state svga_hw_gs_constants = +{ + "hw gs params", + (SVGA_NEW_GS_CONST_BUFFER | + SVGA_NEW_RAST | + SVGA_NEW_GS_VARIANT), + emit_gs_consts +}; -- cgit v1.2.3